3D world generation
Marble →
World Labs' commercial multimodal world model that turns a text prompt, image, video or spatial sketch into an explorable, editable 3D environment, exportable as Gaussian splats and collision meshes. Freemium with paid tiers.
AI agent app
ChatGPT Work →
OpenAI's new agent that merges ChatGPT and Codex for non-technical users, connecting to Slack, Gmail, Drive and CRMs to produce finished documents, spreadsheets, and web apps.
AI app and website builder
Emergent →
A natural-language software-creation platform that generates production-ready websites, apps, and dashboards for non-technical founders and small businesses; just raised a $130M Series C at a $1.5B valuation.
AI assistant / coding agent
Kimi (Kimi K2.6) →
Moonshot AI's web assistant and agent, running the open-weight Kimi K2.6 model; free to use in the browser for chat and long-horizon agent tasks, with the weights also downloadable for self-hosting.
AI assistant / model
Claude Sonnet 5 →
Anthropic's new most-agentic mid-tier model, close to its flagship on hands-on tool and coding work; now the default on Free and Pro plans.
AI coding assistant rules
Hallmark →
A design skill you install into Claude Code, Cursor, or Codex that forces AI-generated interfaces to look designed rather than generated. It picks a different page structure per brief, applies one of 20 named themes, and runs 57 anti-slop gates plus a self-critique pass before emitting anything -- banning fabricated statistics, inline color values, fake browser chrome, and italic headers. MIT licensed, 6.2k stars.
AI coding editor
Cursor →
The AI-native code editor whose real-world developer interaction data trained Grok 4.5; a mature, widely-used tool for agentic coding across many models.
AI compiler / runtime
Modular MAX + Mojo →
A programming language (Mojo) and compiler/runtime (MAX) for running AI models efficiently across different hardware instead of being locked to one chip vendor; now being acquired by Qualcomm but still openly available to developers.
AI gateway / model router
OmniRoute →
A free open-source AI gateway that unifies 230-plus model providers (including many free tiers) behind one endpoint, with token-compression, smart auto-fallback, and multi-agent protocol support, available as a desktop app and PWA.
AI in the browser
Gemma-4 WebGPU Kernels →
A demo running Google's Gemma-4 model directly inside a web browser using your device's graphics hardware — private, on-device AI with no server and no data leaving your machine.
AI security agent
VulnHunter →
Capital One's open-source agentic AI tool that analyzes source code from an attacker's perspective, tries to disprove its own findings before reporting them, and writes targeted fixes. Built for Claude Opus 4.8 in Claude Code; Apache 2.0.
AI security testing
Strix →
Open-source autonomous AI pentesting agents that dynamically find and exploit application vulnerabilities, generate working proof-of-concepts, and integrate with GitHub Actions and CI/CD to block insecure code on every pull request.
AI security tooling
OpenAI Codex Security →
Part of OpenAI's Daybreak program: an agent that builds an editable threat model from your code repository, finds realistic high-impact vulnerabilities, and drafts and tests patches in isolated environments.
AI video production
OpenMontage →
An open-source system that turns an AI coding assistant into an automated video-production studio, with a large library of pipelines, tools, and agent skills for editing and assembling video.
AI-text detection
GPTZero →
The widely used AI-writing detector (about 19M users) that estimates how likely a passage was machine-generated; being acquired by Superhuman to build a persistent authenticity layer.
Agent / automation
Gemini 3.5 Flash computer use →
Google's fast model can now operate a browser, phone, or desktop directly as a built-in tool, with optional confirm-before-acting and auto-stop-on-attack safeguards for building automation agents.
Agent browser
ego-lite →
A macOS browser built so a human and an agent can browse in parallel without fighting over the same window. Ships a substantive browser-automation skill defining a Playwright-like JavaScript surface for agents to drive it.
Agent configuration layer
ECC →
Cross-host configuration for coding agents: shared skills, rules, commands and hooks plus security scans and gates that work across Claude Code, Codex and others, so one policy set follows you between harnesses.
Agent deployment infra
Cloudflare Temporary Accounts →
Lets an automated agent deploy and run on Cloudflare before a human signs up, removing the account-creation step from agent workflows.
Agent experiment framework
harness-training →
A small PyTorch-shaped framework for treating an agent harness as the thing being trained: the harness file is the weights, an improvement agent is the gradient estimator, a deterministic task panel is the loss, and promotion or rejection is the optimizer step. Built around reproducible, deterministic runs.
Agent framework
DeerFlow →
ByteDance's open-source agent harness that breaks a long task into specialist sub-agents running in parallel, executes code safely in sandboxes, keeps memory across sessions, and produces reports, slides, and pages; built on LangChain and works with multiple model providers.
NOOA →
NVIDIA's open agent framework, contributed as the flagship technical artifact of the Open Secure AI Alliance. Its README is candid that it is research software and that its generated-code checks are not a containment boundary, so run agents in OS-level isolation.
Agent infrastructure
AWS Agent Toolkit for AWS →
Official AWS-supported set of MCP servers, skills, and plugins for building AI agents that work with Amazon's cloud services, maintained by AWS itself.
Agent memory
TencentDB Agent Memory →
MIT-licensed memory layer that compresses conversation history into a semantic hierarchy of atoms, scenarios, and personas, with a gateway exposing capture, search, and recall endpoints. Its own benchmarks report token savings in the 31 to 61 percent range. Note that bearer auth and CORS allow-listing both default to off.
Agent memory framework
Microsoft Memora →
Open-source memory system for AI agents that stores rich content but searches it via tiny abstraction labels and cue anchors, cutting token cost on long-horizon tasks. Includes a distillable retriever.
cognee →
An open-source persistent-memory layer for AI agents with remember, recall, forget, and improve operations over a graph-plus-vector store, able to run graph relations, embeddings, session cache, and metadata in a single Postgres instead of four services.
Agent permission control
Claude Code auto mode →
A permission mode that replaces per-action approval prompts with a separate classifier model which blocks escalation, unrecognized infrastructure and actions driven by injected content; becomes the default on 14 August 2026.
Agent platform
Cloudflare OS →
Open-source platform where agents never hold credentials: a Gatekeeper does the OAuth and hands the agent a typed capability scoped to one resource, and every user-built app runs sandboxed with its own SQLite state. Runs locally on workerd for evaluation, or deploys into your own Cloudflare account.
Agent runtime
JarvisHub →
Canvas-native agent runtime where a typed graph of artifacts, versions, dependencies and provenance replaces the chat transcript as the agent's memory and action surface. Ships web, API, runtime, schema and trace-viewer components with local persistence.
Agent safety hook
Destructive Command Guard →
A drop-in hook that blocks catastrophic shell commands (git reset --hard, rm -rf, DROP TABLE) before AI coding agents run them, with sub-millisecond latency and support for nearly every major agent.
Agent sandbox primitive
cloudflare/computer →
MIT-licensed sandboxed filesystem and compute primitive for giving an agent a working machine, which hit number one on GitHub Trending the day it shipped. Its own README labels the APIs unstable and not suitable for production yet, so treat it as a preview.
Agent security middleware
Microsoft Agent Governance Toolkit →
Policy middleware for agent tool calls: binds identity, evaluates policy per action, logs decisions and can deny calls, with Model Context Protocol security checks and prompt-injection detection. Public Preview; app-layer only, so OS isolation still needs containers.
Agent security monitoring
ADR →
Uber's runtime detector for coding agents, watching what agents actually do on developer machines rather than filtering prompts. Reported 206 credential exposures at 97.2 percent precision across 7,200 hosts, and ships with ADR-Bench, a 300-task benign-versus-malicious evaluation set.
Agent security scanner
NVIDIA SkillSpector →
A scanner that inspects agent skills for security problems before you run them -- a static safety check for the fast-growing agent-skill supply chain.
Agent skill
Claude Video →
A /watch skill that downloads a video, extracts adaptive keyframes, pulls existing captions or falls back to Whisper transcription, and hands the material to the host coding agent. An input adapter rather than a planner.
Agent skill pack
reverse-skill →
A deployed cybersecurity skill pack for coding agents: instructions, a routing table that picks the method and tools for a given task type, a local tool inventory, scripts and sub-skills. It also keeps a field journal, writing task outcomes and lessons back to disk so later runs consult prior work. Persistent procedural memory by file mutation, with no verifier checking that each write improves future performance.
Agent skill registry
localskills.sh →
A versioned registry and distribution layer for coding-agent skills. A skill is a folder rooted in SKILL.md with optional scripts, references and assets; versions are immutable and hash-tracked, and installs land in each agent's native location. Its MCP server also lets an agent search and load a skill mid-task, though that copy lives only in the current context window unless installed locally.
Agent skill runtime
Resource2Skill →
Microsoft runtime that compiles tutorials, repos, and articles into structured, executable agent skills with provenance. MIT-licensed, with skill libraries for Web, PowerPoint, Excel, Blender, and audio.
Agent skills collection
mattpocock/skills →
A small, composable set of engineering workflow skills - design review, issue triage, test-driven development, spec generation - deliberately built to plug into your process rather than own it. Installs into any harness that reads the Agent Skills format.
Agent standard
Agentic Resource Discovery (ARD) →
Google's open specification and manifest format (ai-catalog.json) that lets AI agents discover and verify tools and other agents across organizations -- a directory layer for the agent web, backed by Microsoft, Nvidia, Salesforce, GitHub, and Hugging Face.
Agent tooling
chrome-devtools-mcp →
An official MCP server that lets coding agents control and inspect a live Chrome browser, exposing DevTools automation, debugging, and performance analysis to AI assistants.
Agent training toolkit
Skill Self-Play →
Apache-2.0 release of a system that grows and prunes a library of skill packages, each with routing metadata, examples and an executable validator, then trains a solver on the tasks they generate. Includes benchmark material and training launchers; expects eight visible GPUs.
Agentic coding environment
ZCode →
The official desktop harness for Z.ai's GLM-5.2, combining GLM-optimized agents, sub-agents, and long-running 'Goals' with bring-your-own-key model access for planning, coding, review, and deployment.
Agentic data-pipeline platform
DataFlow-WebUI →
An open-source platform where an LLM agent builds persistent, editable data-processing pipelines as validated graphs through a conversational interface and visual editor, instead of emitting throwaway scripts.
Agentic image generation
Meta Muse Image →
Meta's agentic image model that uses test-time compute - searching, coding, and iteratively refining its own output - to reach higher quality than a single-pass generator. (The default Instagram-photo training was pulled after backlash; the model remains.)
Agentic search
T-Search →
An open agentic retriever you can try in the browser - it runs multi-round evidence gathering for questions that need several searches chained together rather than one lookup.
Attention kernels
FlashKDA →
Moonshot AI's MIT-licensed kernel implementation of Kimi Delta Attention, the linear-attention mechanism underneath Kimi K3, published ahead of the model weights themselves. Useful today for anyone building or serving bounded-state attention rather than a growing key-value cache.
Benchmark audit material
Poolside trajectory archive →
Poolside published the full agent trajectories behind its Laguna S 2.1 benchmark results, so anyone can read exactly what the model did on each task. Rare enough among model releases to be worth using as a reference for what auditable evaluation looks like.
Benchmark leaderboard
ARC-AGI verified results leaderboard →
ARC Prize's public result pages list each model's score alongside the exact configuration used - model name, reasoning effort and token limits - plus task replays for ARC-AGI-3 runs. Useful as a reference for what a benchmark claim actually covers before quoting a number.
Artificial Analysis Agentic Index →
A public leaderboard averaging agentic benchmarks that give models shell and web access, including a multi-step banking workflow scored on the resulting database state rather than the model's own summary. Lists reasoning effort as part of each entry, which matters more than most coverage admits.
Bioinformatics data tool
gget →
An open-source command-line and Python tool for querying genomic databases with exact, deterministic lookups. New benchmark work showed wrapping an AI agent around gget's 'virus' module lifted viral-sequence retrieval accuracy from as low as 17% to above 90% -- a concrete template for pairing models with hard tools.
Browser AI agent
Gemma Gem →
A Chrome extension that runs Gemma 4 E2B locally through WebGPU using an ONNX build with 4-bit weights, and gives the resulting agent page-reading, clicking, typing, screenshot and JavaScript tools. Worth knowing before you install: the widely quoted ~500MB is the cached download on disk, and the project's own estimates for GPU and system memory during inference are substantially higher and not benchmarked on real devices.
Build with your own documents
RAGFlow →
An open engine for building AI question-answering over your own files and documents.
Chart language for AI agents
Microsoft Flint →
An open-source visualization language that lets agents describe a chart in JSON and compile it reliably to Vega-Lite, ECharts, or Chart.js, with a Model Context Protocol server for direct tool use.
Code security scanner
Semgrep →
Static-analysis security scanner that finds vulnerability classes like broken access control in real codebases, increasingly paired with AI models in its pipeline. Its public benchmark work this week is also a useful, honest reference for how well current models actually find security bugs.
Code-review context tool
code-review-graph →
Parses a repository with Tree-sitter into a local SQLite graph of code entities and relations, then traces callers, dependents, and tests for a changed file to give an agent a narrow, blast-radius review set via MCP, plus a PR-commenting GitHub Action.
Coding agent
Claude Code →
Anthropic's command-line coding agent that reads a whole codebase, edits files, runs tests and fixes failures on its own; it is the tool behind Anthropic's disclosure that Claude now authors most of its production code.
OpenCode →
An open coding agent shown this week to send a fraction of the fixed token overhead of some rivals, with a stable prompt-cache prefix; works against frontier and local models alike.
Coding agent CLI
Qwen Code →
Alibaba's open-source command-line coding agent, whose 30 July update adds persistent background agents, reusable skills and UI-agent tooling. Free to run against local or hosted Qwen models.
Coding agent harness
Grok Build →
xAI's agentic coding harness and terminal interface, published under Apache 2.0. Genuinely useful if you want to read how a frontier lab wires a production coding agent, and the license lets you fork and ship it. Note the governance: the contributing guide says external contributions are not accepted, and the repo is a one-way bot-pushed mirror of an internal monorepo.
Coding-agent spec format
design.md →
A simple convention from Google Labs for writing a DESIGN.md file that gives an AI coding assistant the context and intent it needs before it starts writing code, aimed at fewer wrong turns on bigger tasks.
Computer-use agent model
Fara 1.5-27B →
Microsoft's MIT-licensed 27B multimodal agent that operates web browsers from screenshots alone, emitting clicks, typing, scrolling, and navigation, and trained to pause on ambiguity or unauthorized irreversible actions.
Consumer health assistant
Health in ChatGPT →
OpenAI's health surface, rolling out to US adults on web and iOS since July 23. With permission it connects Apple Health data and medical records, then uses that context inside ordinary ChatGPT conversations to help compare results and prepare for appointments. OpenAI stresses it supports rather than replaces clinicians.
Cost efficiency
Caveman →
A skill that compresses AI agent responses into terse output, cutting roughly 65% of output tokens while preserving technical accuracy across 30-plus coding agents like Claude Code, Cursor, and Gemini.
Create images & video
ComfyUI →
A visual, node-based studio for generating images and video with open models. Powerful and endlessly extensible.
Cut AI agent costs
Headroom →
A drop-in proxy that sits between your coding assistant and the AI model and automatically compresses bulky tool outputs, logs, and retrieved text before they reach the model — cutting token usage sharply without changing your code.
Dataset opt-out checker
Am I in The Stack? →
Lets a developer check whether their GitHub repositories were included in The Stack code dataset, and points to BigCode's removal process. Opted-out repositories are dropped before each patch release.
Deployment guide
vLLM recipe for DeepSeek V4 Flash →
An official vLLM recipe page with working launch commands for serving V4 Flash across several hardware configurations, including the flag that turns on the DSpark speculative-decoding module and the FP8 KV-cache and expert-parallel settings DeepSeek recommends.
Diffusion LLM API
Mercury 2 (Inception Labs) →
An API-only diffusion language model pitched on raw speed, claiming to out-pace open diffusion models on tokens-per-second for latency-sensitive generation.
Distributed inference
Mesh LLM →
Runs models too big for one machine by splitting them across networked peers over serverless peer-to-peer transport; ~18 MB install, 40+ models up to 235B, OpenAI-compatible API on localhost.
Document OCR
Unlimited OCR →
Baidu's 3-billion-parameter document parser transcribes dozens of pages in a single pass without its memory footprint growing, because its decoder holds a constant-size cache instead of one that expands with every token. MIT licensed, with vLLM, ModelScope and ms-swift support already wired in, plus a hosted demo you can try in a browser.
Document reading (hosted)
Mistral OCR 4 →
A hosted document-reading model that converts scanned pages, PDFs, and complex layouts into clean structured text ready for a language model. Send a document, get back tidy text with the structure preserved.
Document-to-text for AI
MinerU →
Open-source tool that converts complex PDFs and office files into clean markdown and structured data that AI models can read reliably. Run it yourself for free, with nothing leaving your machine.
E-ink app
riddle →
An open-source Rust app that turns a reMarkable Paper Pro into an interactive AI diary - handwrite a question and a vision LLM writes back in animated e-ink handwriting. Works with any OpenAI-compatible API. MIT-licensed.
Efficient open model
Nemotron-3-Puzzle-75B →
Nvidia's compressed 75B open model (from a 120B parent) with roughly double the serving throughput and 8x long-context concurrency on a single H100; weights on Hugging Face.
Embedding model
NVIDIA Nemotron 3 Embed 8B →
8-billion-parameter retrieval encoder that turns queries and documents into normalized dense vectors for semantic search. NVIDIA claims state-of-the-art results on the multilingual RTEB leaderboard as of July 16; released under OpenMDW 1.1.
Enterprise agent platform
Claude Tag (agent identity access model) →
Anthropic's product for putting Claude to work in shared team channels, now with an access model that gives each agent its own scoped accounts in the systems it touches -- GitHub, Slack, a data warehouse -- instead of borrowing an individual user's permissions, so every action is bounded and audited.
Experimental inference engine
WASTE →
A C inference engine that streams only the experts a mixture-of-experts model actually activates directly off NVMe, using spare RAM as an expert cache. Its author reports running the full 2.78-trillion-parameter Kimi K3 on a 64 GB laptop at about half a token per second. An existence proof, not a chat app.
Fast agentic LLM (API)
Gemini 3.6 Flash →
Google's newly GA fast model streams output nearly twice as fast as 3.5 Flash and costs less per task while holding the same intelligence-index score, tuned for high-volume agent loops that use fewer tokens and tool calls.
Find models & datasets
Hugging Face →
The main hub for finding, downloading, and trying open AI models and datasets — the field's town square.
Fine-tuning API
Tinker →
Thinking Machines Lab's hosted fine-tuning service, now serving Inkling alongside its other models. It is the managed path to customizing Inkling if you do not want to provision the GPUs yourself -- with the caveat that the API caps context at 256K tokens, versus 1M for the open weights you run yourself.
Fine-tuning toolkit
Unsloth (AMD support) →
The fine-tuning and RL toolkit now documents AMD support across training, RL, chat, and deployment on Windows, WSL, and Linux, plus a cross-platform Studio beta.
Forecasting model
TimesFM →
Google's pre-trained foundation model for time-series forecasting — predicting things that change over time, like demand, traffic, or sensor readings — usable out of the box without training your own model.
Formal mathematics library
Formal Conjectures →
Google DeepMind's open Lean library of formally stated open mathematical conjectures, now the venue where the claimed Jacobian conjecture counterexample is being reviewed in public. A usable resource if you want machine-checkable statements of open problems rather than prose.
Frontier chat & coding model
Grok 4.5 →
SpaceXAI's new 1.5-trillion-parameter model, available in Grok Build, Cursor, and the API at $2 per million input / $6 per million output tokens, with a full public release on July 9.
Frontier model
Claude Fable 5 (redeployed) →
Anthropic's top-tier model, back online after a brief export-control suspension, now shipping with a hardened cybersecurity classifier that reroutes flagged requests to Opus 4.8 and a wider default safety margin.
Frontier model API
GPT-5.6 (Sol / Terra / Luna) →
OpenAI's newest model family, tuned for cheap, fast, reliable agentic work, with programmatic tool calling, a multi-agent beta, persisted reasoning, and a high-reliability 'pro' mode.
GPU kernel verification
FpSan (Floating-Point Sanitizer) →
Open-source correctness checker for Triton GPU kernels, and the tool OpenAI says it used to validate the production kernels GPT-5.6 Sol rewrote. It compares symbolic computation under its own payload algebra rather than simulating IEEE floating point, so results should be compared only against other FpSan runs. Useful for anyone writing or generating custom kernels who needs to catch numerical breakage before it reaches production.
Genome language model
Evo →
Arc Institute's family of genome language models, released openly with code and checkpoints. Used by Arc and Stanford to generate complete synthetic bacteriophage genomes that were then built and tested in the lab against non-pathogenic bacterial hosts.
Genome model
Evo 2 →
Arc Institute's open genome language model for DNA, used to design bacteriophage genomes that were synthesized and shown to work in living bacteria; weights and code are public.
Give AI agents code memory
codebase-memory-mcp →
Indexes an entire codebase into a persistent, queryable knowledge graph so AI agents can understand large projects fast. Supports a huge range of programming languages, answers queries near-instantly, and ships as a single dependency-free binary.
Hardware peripheral
Codex Micro →
A $230 mechanical control deck for driving OpenAI's Codex agents, built with keyboard maker Work Louder. 13 switches, a joystick, a touch sensor, RGB keys showing live agent status, and a rotary dial that adjusts reasoning effort -- turning an API parameter into a physical knob. Nothing it does is impossible with keyboard shortcuts; the pitch is ambient awareness when supervising several agents at once.
Hosted LLM API
DeepSeek V4 Pro (API) →
A strong open-weight reasoning and coding model now offered through DeepSeek's own API at a permanently cut, low per-token price, undercutting frontier closed models for high-volume work.
Hosted demo
Galahad verified-reuse testbench →
Public testbench for the frozen-12B verified procedure cache, where a solved and independently verified problem family is answered on later instances at zero generation tokens, bit-exact. Worth poking at to understand what the claim does and does not cover -- the engine source, configuration and raw artifacts are withheld, so this demo plus the bench repo is the only inspectable surface.
Hosted model API
DeepSeek V4 Flash 0731 (API) →
The updated V4 Flash checkpoint now serves behind the existing deepseek-v4-flash identifier, with a 1-million-token context, 384K maximum output, tool calls, and an OpenAI-, Anthropic- and Responses-API-compatible interface. Fresh input runs $0.14 per million tokens, output $0.28, and cached input $0.0028 - a fiftyfold discount on repeated prefixes.
Gemini 3.6 Flash and 3.5 Flash-Lite →
Google's economy-tier models went generally available on July 21, with 3.6 Flash keeping a million-token context and 64,000-token output while dropping its output price roughly a sixth versus 3.5 Flash and using about 17 percent fewer output tokens per task. Note the migration-breaking changes: some sampling parameters are deprecated and prefilled model turns are no longer supported.
Qwen3.8-Max →
Alibaba's new flagship multimodal model, live today as a paid API at $2 per million input tokens and $6 per million output tokens, with a one-million-token context, function calling, structured output, and prompt caching that drops repeated input to $0.25 per million. Weights are promised but not published.
Seed2.0 (ByteDance Seed) →
ByteDance's Seed2.0 family (Pro, Lite, Mini) of closed, API-hosted models aimed at long-tail knowledge and complex instruction-following, accessed through ByteDance's Volcano Engine (Ark) platform. Not open weights despite the academic-style model card.
Hosted open-model API
GLM-5.2 on Baseten →
The top trending open-weight model served as a fast hosted endpoint, reported at 280+ tokens/sec on Blackwell-class hardware -- an open model you can call like a closed one.
Image generation
Mage-Flow →
A Microsoft demo space for image generation and editing that works at native resolution rather than upscaling from a fixed square, running free on Hugging Face's shared GPU tier.
Muse Image →
Meta's agentic image model, free for everyday creation inside Meta AI, Instagram Stories (US), and WhatsApp; it can search, write code, and self-refine rather than mapping a prompt straight to pixels, and stamps outputs with an invisible Content Seal watermark.
Image generation API
Nano Banana 2 Lite →
Google's fastest, cheapest Gemini image model - a text-to-image picture in about four seconds for roughly three cents per thousand images, built for high-volume use.
Image generation and editing
Boogu-Image 0.1 →
An open-source unified image understanding and generation model family (Base, Turbo, Edit, Edit-Turbo) with instruction-based editing and bilingual Chinese-English text rendering, trained for roughly $400K. Apache 2.0.
Image, video and audio generation
FLUX 3 (early access) →
Black Forest Labs' unified generation model, producing video up to 20 seconds with native synchronized audio from text, image, video or keyframe inputs. Video is behind an early-access request today; image access is promised in the following weeks and open weights are deferred.
Inference acceleration
Sol-Attn (Sol-Engine) →
NVIDIA's drop-in sparse attention kernel for long-video diffusion transformers, released July 28 for HunyuanVideo-13B and Wan2.1-T2V-14B. Screens compressed key/value blocks inside a single online-softmax pass, so exact attention goes where it matters and skipped blocks get an approximate correction. Training-free, no weight changes, reported up to 2.1x for generation and 2.3x for editing. The repo marks end-to-end re-benchmarks for the two integrated pipelines as pending.
Inference engine
SGLang (Kimi K3 cookbook) →
Alternative open-source serving engine with day-zero K3 support and a step-by-step deployment cookbook. Its writeup documents how prefix caching, paging and prefill/decode disaggregation were rebuilt to handle K3's mix of recurrent and key-value state.
Inference platform
Doubleword (async + batch inference) →
Run the same models you already use, but on async and batch tiers that trade latency for a large cost cut on workloads that don't need an instant reply: long-running agents, evaluations, and bulk jobs.
Inference runtime
LvLLM →
A community inference runtime specialised in hybrid CPU and GPU execution of mixture-of-experts models, with NUMA-aware scheduling, expert weight management, and MXFP4 quantization kernels. Its DeepSeek V4 build publishes a working dual-RTX-3090 launch configuration at 22K context.
Inference server
vLLM (Kimi K3 support) →
The widely used open-source serving engine landed day-zero Kimi K3 support with a documented recipe, an FAQ on minimum hardware, and a K3-specific DSpark draft model for speculative decoding that roughly triples single-user throughput.
vLLM DeepSeek-V4 support →
vLLM shipped serving support for DeepSeek-V4's compressed long-context attention, including hybrid KV-cache management, multiple cache page sizes, kernel fusion and multi-stream partitioning. The engineering post documents the recipe and the hardware it assumes.
Interactive demo
Neuronpedia J-lens demo →
A live, no-install web demo of the Jacobian lens that lets you watch the 'contents of the workspace' light up inside open models (Qwen 3.6 27B and Gemma 3 12B) as they process text.
Interpretability tool
Jacobian Lens (J-lens) →
Anthropic's open-source tool that reads a model's silent 'working memory' - for any word, it finds the internal pattern that makes the model more likely to say it later. Apache-2.0, with a live interactive demo on open models.
jlens-gguf →
A GGUF-native implementation of Anthropic's Jacobian Lens for local models, with a browser UI to visualize, swap, and ablate a model's internal concepts live as it generates through llama.cpp.
LLM API and open weights
DeepSeek V4 →
DeepSeek's latest model family (a 1.6T-parameter Pro and a 284B Flash, both with a 1-million-token context by default), available as an API and as open weights on Hugging Face.
LLM API router
Dahl Inference →
Third-party inference router reselling top open-weight models (Kimi K2.6, MiniMax M2.7, GLM 5.2) at low per-token prices, currently running a 100M-free-token promotion.
LLM chat and API
Kimi K3 →
Moonshot AI's 2.8-trillion-parameter flagship with a 1M-token context window, tuned for agentic coding and knowledge work; it topped a frontend-coding leaderboard. Usable now via kimi.com chat and an OpenAI-compatible API, with open weights due July 27.
LLM cost routing
World Model Optimizer →
A pip-installable CLI that turns the OpenTelemetry traces your agents already emit into a routing policy: it scores every model you have registered against held-out tasks from your own traffic, then serves an endpoint that sends easy requests to cheap models. Treat the routing as the product; the distillation half has no released checkpoint yet.
LLM inference acceleration
DeepSeek DSpark →
Open-source speculative-decoding implementation using parallel tree drafting to speed up text generation with no change to the model's output - the project that topped Hacker News this week. Drop-in inference speedups for self-hosted models.
JetSpec →
Parallel tree-drafting speculative decoding aiming for large, lossless inference speedups; project page and writeup with code, reporting up to several-times faster generation depending on the model and workload.
Language model API
Ling-3.0-flash (free API) →
Ant's 124B-parameter mixture-of-experts model that activates only 5.1B parameters per token, with a 256K context and OpenAI- and Anthropic-compatible endpoints. Currently free on OpenRouter as inclusionai/ling-3.0-flash:free; aimed at long-horizon agent workflows and tool calling.
Local AI runtime
Ollama 0.31 →
Run open models on your own computer; the new version nearly doubles Gemma's speed on Apple Silicon using multi-token prediction, on by default.
Local coding model
Gemma-4 12B Coder (GGUF) →
A fine-tuned, locally-runnable version of Google's Gemma-4 model specialized for programming tasks, packaged in a format that runs efficiently on everyday consumer hardware.
Local coding-agent model
Poolside Laguna S 2.1 →
A public-weight, 118B-total mixture-of-experts coding model with only ~8B active parameters that runs locally on a single 128GB machine via a 75GB Q4 GGUF, built for long-horizon agentic software work under the permissive OpenMDW-1.1 license.
Local inference
TurboFieldfare →
A Swift and Metal runtime that runs Gemma 4's 26B model on an 8GB MacBook Air by keeping a 1.35GB core resident and streaming the rest of the experts off the SSD. Ships as a Mac app, a CLI and an OpenAI-compatible local server.
Local inference engine
bitnet.cpp →
Microsoft's official inference framework for 1.58-bit ternary language models, built on llama.cpp with optimized CPU and GPU kernels for running very heavily compressed models on ordinary hardware.
llama.cpp b10228 →
The release that adds DeepSeek V4 Flash's embedded DSpark speculative-decoding head, plus a converter that can split the draft tensors into a separate GGUF. Gains are workload-dependent: roughly 2x decode on large multi-GPU setups, and a measured slowdown on a 24 GB card with CPU offload.
Local inference runtime
CachyLLama →
MIT-licensed llama.cpp fork that saves conversation and system-prompt caches to SSD and restores them after a restart, so local agents stop reprocessing the same prompt prefix every turn. Its own benchmark reports long repeated agent prefixes going from minutes cold to about a second warm.
FastFlowLM →
An NPU-first, GPU-free inference runtime built exclusively for AMD Ryzen AI (XDNA) NPUs, targeting long-context local LLMs at low power on laptop-class hardware; the team just joined AMD, with open install guides for Ubuntu, Arch, and more.
llama.cpp b10217 →
The 1 August build adds support for DeepSeek V4 Flash emitting tool calls inside its reasoning block, which is what was silently killing local agent runs against the new model. If you are running DS4 locally with tools, this is the build you need.
Local inference server
llama.cpp (MCP tool hosting) →
The most widely used local LLM server now launches and manages local Model Context Protocol tool processes itself, discovers their tools and exposes them through its chat API - turning a plain inference server into an agent host. Off by default; needs a tool-capable chat template.
Local model package
Maple-Preview (2-bit MLX build) →
DeepGrove's 20B mixture-of-experts model with about 1B active parameters per token, packaged for Apple Silicon at roughly 5.3GB. The build uses affine two-bit group quantisation with four-bit embeddings and output head, and its loader packs ternary values into two-bit codes. Note that the native BF16 repository is about 40.4GB, and DeepGrove publishes no ternary training recipe or independent evaluation.
Local model runner / fine-tuning
Unsloth →
Toolkit and documentation for running and fine-tuning large open models faster and on smaller hardware, including aggressive dynamic quantization recipes that shrink models like GLM 5.2 by 80-plus percent while keeping most of their accuracy. The practical on-ramp to running near-frontier models privately.
Local model runtime
SwiftLM →
An MLX-based runtime for Apple silicon whose --stream-experts mode reads mixture-of-experts weights straight off an NVMe SSD, letting a machine run models several times larger than its RAM.
Local model server
KoboldCpp v1.118 →
Single-binary local model server that shipped its own fix for multi-turn DeepSeek V4 Flash prompt-processing problems on the same day as the llama.cpp fix. Useful if you want a working DS4 setup without building anything.
Local model weights
DeepSeek-V4-Flash-0731 GGUF (Unsloth) →
Community quantizations of the new MIT-licensed DeepSeek weights in GGUF form, running from roughly 83GB at aggressive low precision to about 162GB at 8-bit. Usable on high-memory workstations and multi-GPU rigs, not on a laptop.
Local voice AI toolkit
Voicebox →
A local-first voice stack bundling voice cloning, TTS, Whisper transcription and dictation, a refinement model, a REST API, and a built-in MCP server so an agent can speak, transcribe, and manage voice profiles without cloud calls.
Low-bit language model
Bonsai 27B (GGUF) →
PrismML's roughly 27.8-billion-parameter Qwen-derived model trained with 1-bit binary or 1.58-bit ternary weights end to end, which the company says fits in about 4 GB and runs on phone-class hardware. Performance figures are vendor-reported and not independently replicated.
MCP app framework
Skybridge →
A framework for building MCP-native apps -- interactive tools an AI assistant can open and use directly, pitched as 'MCP apps are the new website.'
MCP server/client toolkit
FastMCP →
A Python toolkit that turns ordinary functions into Model Context Protocol tools, resources, and prompts with generated schemas, validation, and docs, and a client that handles transport negotiation, auth, and protocol lifecycle.
Model API
Meta Model API (Muse Spark 1.1) →
Meta's first paid, hosted model API, built around the Muse Spark 1.1 multimodal reasoning model -- a million-token context window with active context compaction, zero-shot tool and MCP support, and an OpenAI-compatible interface so existing code drops in with little more than an endpoint change.
Model comparison
Artificial Analysis Intelligence Index →
The independent benchmark and pricing dashboard the field now reaches for when a lab claims a lead -- it is the source of Inkling's debut score of 41. Useful beyond the headline ranking because it also tracks output tokens per task, latency and cost, which is how you find out that a cheaper-looking model is actually more expensive per finished job.
Model comparison and benchmarking
Artificial Analysis →
A free public dashboard that independently benchmarks and compares AI models on a combined intelligence index alongside price and speed; the source of this week's finding that GLM-5.2 leads the open-weight class.
Model control heads
Multi-Head Latent Control →
Freezes a model and attaches two small heads that read its hidden states to decide whether to answer, use a tool, ask for information, abstain, or escalate to a stronger model. Open-sourced with matching small checkpoints; needs white-box access.
Model gateway
Vercel AI Gateway (Ling-3.0-flash) →
Vercel added Ling-3.0-flash to its AI Gateway with bring-your-own-key support and failover routing, free through August 3. Useful if you want the model behind a single gateway alongside other providers rather than wiring a second API.
Model integrity auditing
abliterlitics →
An evaluation harness for checking whether an edited or guardrail-stripped model is actually intact: it diffs every tensor against the base model, measures behavioural drift on harmless prompts, runs a multi-domain capability suite, and scores harmful-completion rates separately. A tensor diff from this would have caught this week's broken Gemma 4 release in seconds.
Model quantization
Unsloth Kimi-K3-GGUF →
Converted local-inference builds of Moonshot's Kimi K3: a 1.51 TB four-bit UD-Q4_K_XL file, a 1.56 TB eight-bit build, and BF16/F16/F32 multimodal projector files that preserve an image-input path. Datacenter-scale hardware still required.
Voodoo Quant →
A per-tensor sensitivity-aware quantization method that spends more bits on important tensors, claiming large divergence reductions over standard llama.cpp and Unsloth quants, especially at 1-bit and 2-bit; GGUF files run in unmodified llama.cpp.
Model serving engine
SGLang →
The other serving stack DeepSeek's official 0731 model card documents as supporting DSpark directly, alongside the recommended FP8 key-value cache and FP4 indexer cache configuration for V4 Flash.
Model weights
Inkling-Small GGUF →
Quantized builds of Thinking Machines' newly released 276B/12B multimodal open-weight model, packaged for llama.cpp, LM Studio and Ollama so you do not have to download the 532GB original.
Model-orchestration API
Sakana Fugu →
A single OpenAI-compatible endpoint that dynamically routes each request across several frontier models, so you call one API and get a coordinated multi-model answer.
Multi-agent framework
agency-agents →
An open-source library of 150-plus specialized AI agent personas across 13-plus professional divisions, built to run multi-agent workflows natively in Claude Code with conversion scripts for other agentic coding tools.
Multi-agent research system
SearchOS →
Open-source (MIT) multi-agent web-research framework that treats search like an operating system: progress lives in an explicit evidence graph, coverage map, frontier task queue, and failure memory instead of chat history, with a pipeline-parallel scheduler. Ships a CLI/TUI, web frontend, installer, and replayable sessions.
Office-document access for agents
OfficeCLI →
A command-line tool that lets AI agents read and edit Word, Excel, and PowerPoint files, one of the week's fastest-rising agent-infrastructure repos on GitHub.
On-device AI
Program-as-Weights demo →
A public demo and code for compiling natural-language task specs into tiny neural artifacts that run locally on a frozen small model, matching much larger models on narrow fuzzy tasks.
On-device LLM runtime
Cactus Hybrid (Gemma-4 E2B) →
A phone-sized Gemma-4 checkpoint with an attached error probe that scores how likely each answer is wrong and routes low-confidence queries to a cloud model; weights and runtime are public (set CACTUS_CLOUD_STRICT_SSL before using the cloud path).
On-device language model
Bonsai 27B →
PrismML's 1-bit and 1.58-bit builds of Qwen3.6 27B, compressing a 54 GB model to 3.9 GB (binary) or 5.9 GB (ternary) and running at roughly 11 tokens per second on an iPhone 17 Pro. The release ships an honest benchmark table showing the cost: instruction following, tool calling, and vision all degrade sharply, and the vendor states agentic coding is not a strong target of this release.
On-device model
MiniCPM5-1B →
OpenBMB's dense 1B local model with Think and No-Think modes, trained with SFT, RL, and on-policy distillation. Designed for on-device and edge deployment.
On-device model compiler
Program-as-Weights →
Turns a plain-English task spec into a small weight file that a frozen 0.6B model runs locally -- matching a 32B model's quality at roughly one-fiftieth the memory and about 30 tokens/sec on a MacBook M3. Open repo and site for compiling cheap, offline 'fuzzy' text programs.
On-device model runtime
Noema Overfit →
Repackages compatible mixture-of-experts model files so shared weights stay resident in memory while expert weights stream from local storage on demand, letting phones load models far larger than their RAM. Experimental, and slower than a smaller fully-resident model on short prompts.
On-device quantized model
Ternary-Bonsai-27B (GGUF) →
PrismML's ternary-weight 27B model in GGUF at ~7.2 GB deployed, with custom CUDA/Metal/CPU kernels. Expands local hardware reach, though agentic reliability is still limited per early tests.
On-device speech recognition
Apple SpeechAnalyzer →
Apple's on-device speech-to-text API that cut errors roughly fourfold over the legacy recognizer and beat Whisper Small using about a third of the compute - private, local transcription with no cloud round-trip.
On-device world model
Cosmos3-Edge →
NVIDIA's compact 4-billion-parameter physical-AI model generates text autoregressively while producing image, video, audio and action-trajectory outputs through a diffusion tower, sized for local robotics, autonomous-vehicle and smart-infrastructure workloads. NVIDIA warns it is not physically accurate simulation or safety-certified reasoning.
Open LLM
Qwen3.6 (open weights) →
Alibaba's stable Qwen3.6 release: open-weight general chat and coding models you can self-host, the same family at the center of this week's open-vs-closed pricing debate.
Open LLM weights
Xiaomi MiMo-V2.5-DFlash →
Xiaomi's official DFlash release on Hugging Face -- a 1-trillion-parameter mixture-of-experts model (42B active) under an MIT license, with FP4 quantization and parallel decoding for high inference throughput.
Open coding model
LongCat-2.0 →
Meituan's 1.6T-parameter MoE model tuned for coding and agentic work, MIT-licensed weights plus a cheap hosted API (launch promo $0.30/$1.20 per million tokens) that self-hosts to avoid data-jurisdiction concerns.
Open image model
Qwen-Image-2.0-Pro →
Alibaba's latest open image-generation model in the Qwen family, downloadable and runnable locally, part of a broad open-weight release wave that also refreshed the Qwen3.6 chat models.
Open language model
LLaDA / iLLaDA →
An openly released diffusion language model (weights and code) that generates text by refining a whole passage at once rather than one word at a time, useful for experimenting with non-autoregressive generation and infilling.
Open large language model
GLM-5.2 →
A flagship openly-available language model with a very large context window for long documents and code. Free to download and run yourself, with compressed versions for more modest hardware.
Open model
GLM 5.2 (GGUF, runnable locally) →
Zhipu AI's open, MIT-licensed mixture-of-experts model with a roughly million-token context, now packaged as ready-to-run quantized files you can host on your own machine. Strong on agent and coding workflows; this week it beat Claude on a narrow security benchmark at a fraction of the cost.
Open model download
Kimi K2.6 weights (Hugging Face) →
The actual Kimi K2.6 model weights, published under a modified-MIT license for anyone to download, run, and build on; large enough that full-strength use needs a multi-GPU node.
Open model for formal math
Leanstral 1.5 →
A free, open mixture-of-experts model specialized for writing machine-checked Lean 4 proofs and translating ordinary math into formal, verifiable form.
Open model weights
Frontis-MA1-35B →
A 35-billion-parameter open model post-trained specifically to write, run, debug and recombine machine-learning code inside an evolutionary search loop. Released with the full OpenMLE stack, so the search framework it was trained for is public too.
Open robotics model
RxBrain (Hy-Embodied-RxBrain-1.0) →
Tencent's ~6.2B embodied model that interleaves text reasoning with generated goal images to plan robot tasks. Weights and inference code released under Apache-2.0.
Open video understanding model
VideoChat3-4B →
A fully open 4B-parameter video multimodal model for general, long-form, and streaming video understanding, released with weights, training code, training strategy, and datasets.
Open-source agent framework
OpenClaw →
The fastest-growing repo on GitHub, now a MIT-licensed nonprofit, a neutral open framework for building AI agents that plug into any model or lab.
Open-source speech recognition
OpenAI Whisper →
OpenAI's open-source speech-recognition model family and the reference baseline Apple's SpeechAnalyzer was measured against - freely runnable locally in sizes from tiny to large for transcription and translation.
Open-weight LLM
Qwen3-Next-80B-A3B-Instruct →
Alibaba's efficiency-focused open-weight model (80B total / 3B active, 512 experts) with 262K native context to ~1M, built around hybrid attention, high-sparsity MoE, and multi-token prediction; the model card claims roughly 10x inference throughput past 32K context versus a dense 32B baseline.
Qwen3.6-35B-A3B →
Alibaba's open-weight agentic-coding model (35B total / 3B active, Apache 2.0) with 262K native context extensible toward 1M tokens, hybrid Gated-DeltaNet + MoE attention, thinking preservation across turns, and built-in tool use. Downloadable and self-hostable on common open serving stacks.
Open-weight agent model
Solar Open 2 →
Upstage's 250-billion-parameter mixture-of-experts model activates only 15 billion parameters per token and runs on two NVIDIA H200 GPUs once quantized, with a one-million-token context aimed at long multi-step agent work. Weights and a full technical report are public under a custom license requiring Solar-prefixed derivative names and Built with Solar attribution.
Open-weight coding model
Kimi K2.7 Code →
Moonshot AI's trillion-parameter mixture-of-experts coding agent, with only 32B active per token, a 256K context, and vision input, now selectable inside GitHub Copilot and downloadable under a Modified MIT license.
Poolside Laguna S 2.1 (GGUF) →
Open-weight 118B mixture-of-experts coding agent activating about 8B parameters per token, under the permissive OpenMDW-1.1 licence, in GGUF plus FP8, NVFP4 and INT4 builds. Use the current re-released Q4/Q8 files - the initial ones shipped with a broken chat template.
Open-weight language model
Gemma 4 26B A4B →
Google's compute-efficient multimodal model with 25.2 billion total parameters but only 3.8 billion active per token, aimed at running usefully on hardware that cannot host a dense model of comparable capability.
Open-weight language models
DeepSeek-V4 (Pro & Flash) →
Two newly previewed open-weight models with a 1-million-token context window on by default - a large mixture-of-experts flagship and a smaller, fast everyday model. Downloadable weights plus an API.
Google Gemma (open weights) →
Google's open-weight model family, light enough that developers are now embedding it directly into interactive apps - including a demo running Gemma inside the Godot game engine via Vulkan compute shaders, no Python server required.
Open-weight model
DeepSeek V4 Flash 0731 →
The current V4 Flash checkpoint, with weights, the DSpark draft head embedded, and the encoder file that reveals the reasoning-effort labels are prompt prefixes rather than a compute dial. The card also specifies the intended FP8 key-value cache and FP4 indexer cache serving recipe.
DeepSeek-V4-Flash →
MIT-licensed weights for DeepSeek's 284B-total / 13B-active mixture-of-experts model with a one-million-token context, with vLLM and SGLang serving examples on the model card. Real hardware bar: the reference recipe targets four B200 or B300 GPUs.
Ling-3.0-flash →
inclusionAI's 124B mixture-of-experts model with about 5.1B parameters activated per token. Sparse routing genuinely cuts per-token compute, but this is a server-class artifact, not a laptop one: the BF16 repository is roughly 255GB and the official serving path calls for custom SGLang or vLLM forks with tensor parallelism across four GPUs.
MiniMax-M3 →
A natively multimodal open model trained on text, image, and video from the first step, with a million-token context and a sparse-attention design built for speed; downloadable for self-hosting and also offered through MiniMax's own API and agent platform.
Open-weight model (agent world model)
Qwen-AgentWorld →
Alibaba's open language world model that simulates agent environments -- browser, terminal, phone, coding workspace and more -- so other agents can be trained inside the simulation. Released with open weights and code in two sizes.
Open-weight model (self-host)
DiffusionGemma →
Google's open-weight text-diffusion model that generates text in parallel blocks instead of one token at a time; Apache-2.0, runnable locally, with community tooling already shipping.
Open-weight multimodal models
Gemma 4 →
Google's downloadable model family (2.3B-31B, dense and MoE) that natively handles text, vision, and audio, including a 12B encoder-free variant and a thinking mode.
Open-weight reasoning model
Ring-2.6-1T →
Ant Group's trillion-parameter mixture-of-experts reasoning model, activating roughly 63 billion parameters per token, with 128K context extendable to 256K. All checkpoints openly downloadable under the MIT license, with high and xhigh reasoning-effort settings that trade depth against speed and cost. Benchmark claims are vendor-supplied and measured against a previous generation of rivals.
Open-weight research agent
AREX-Turbo →
Apache-2.0 4-billion-parameter deep-research agent from BAAI that audits its own provisional answers against the question's constraints and re-runs research when confidence is low. The public quick-start exposes search and page-visit tools but not the paper's full outer control loop.
Open-weights model
Inkling →
Thinking Machines Lab's 975B-parameter mixture-of-experts model, released July 15 under Apache 2.0. Only ~41B parameters activate per token, it accepts text, image and audio input, and it handles up to 1M tokens of context. Artificial Analysis ranks it the top US open-weights model. Free to download, modify and use commercially -- but you will need serious hardware to run it.
Post-training toolkit
SLAI T-Rex →
The public workflow behind a full-parameter Ascend post-training run on a DeepSeek-V4-family model: FP8-to-BF16-to-Megatron checkpoint conversion, launch templates, and inspectable data-construction pipelines for continued pre-training and supervised fine-tuning. The production engine and custom kernels are withheld.
Public data API
Stack Exchange API →
The free, key-less API behind Stack Overflow and its sister sites, which will return exact question, answer and user counts for any date range -- useful for checking claims about the site's decline yourself.
Quantised model weights
Unsloth DeepSeek-V4-Flash-0731 GGUF →
Published quantisations of DeepSeek's 671-billion-parameter Flash model, ranging from roughly 91 GB at two bits to 162 GB at eight. The card is also the clearest available statement of what hardware each tier actually needs.
Quantization toolkit
TurboQuant-MLX →
Quantization tooling for MLX with published size and speed measurements, including a 3-bit path that takes a 120-billion-parameter model from about 63GB down to 48GB on consumer Macs.
Real-time interactive video
Wan Streamer v0.3 →
Streaming audio-visual interaction model that treats a video as a persistent world plus a time-varying event stream, running full-duplex real-time conversation at 640x368 / 25fps with roughly 550ms total interaction latency.
Real-time voice
GPT-Live →
OpenAI's full-duplex voice interface that talks, listens, and interrupts in real time while delegating deep reasoning to GPT-5.5 in the background; free mini tier plus a paid tier.
Reasoning assistant + API
Muse Spark 1.1 →
Meta Superintelligence Labs' multimodal reasoning model built for agentic work - tool and computer use, coding, a 1M-token context window, and subagent orchestration; live in the Meta AI app's Thinking mode and on meta.ai, with a Meta Model API in public preview.
Regulatory compliance reference
AI Act Service Desk →
The European Commission's official article-by-article guide to the AI Act, including Article 50 transparency duties that apply from 2 August 2026. Note its own warning that displayed text may lag the latest amendments - check the Official Journal for dates.
Reinforcement-learning training framework
prime-rl →
Open-source RL post-training stack that splits rollout generation and gradient updates across GPUs, used in this week's widely discussed $500 fine-tune that beat five frontier configurations on a catalog-review workflow. Practical for teams that already have an automatically scored task and want to train a specialist rather than pay per call for a frontier model.
Research workbench
Claude Science →
An AI workbench that unifies literature search, notebooks, statistics, and cluster compute, and keeps a reproducible record behind every figure. Beta on Mac and Linux.
Retrieval library
PyLate →
A training and retrieval library for late-interaction models, built on Sentence Transformers, for people who want to fine-tune a retriever on their own corpus rather than use an off-the-shelf embedding API.
Retrieval model
DenseOn →
A fully open 149-million-parameter dense retrieval model from LightOn for multilingual, long-context and code search, released with its training data and training code rather than weights alone.
LateOn →
LightOn's late-interaction counterpart to DenseOn - it keeps a vector per token instead of one per document, which costs more storage but retrieves noticeably better on hard queries.
Reward model
OS-Shepherd-9B →
A 9B reward model trained specifically to judge whether a computer-use agent actually finished its task, built to cut the false-success verdicts that general-purpose vision-language judges produce. A 35B sibling and the OSReward benchmark ship alongside it.
Robot manipulation dataset
HiFi-UMI-2K →
Released dataset behind this week's handheld-only robot training result: high-fidelity two-handed human demonstrations captured with a head-mounted stereo rig and tracked grippers, with every trajectory reconstructed and rejected unless a target robot could physically replay it. Covers wiping, shirt folding, remote insertion and produce sorting. Directly usable for imitation-learning experiments without owning a teleoperation setup.
Robot navigation model
Robostral Navigate →
Mistral's 8B embodied model that steers wheeled, legged, or flying robots through unseen environments from a single RGB camera and a plain-language instruction.
Robot policy model
WorldDiT →
Four released checkpoints plus self-contained inference and evaluation code for a sub-billion-parameter diffusion transformer that emits continuous robot action chunks while predicting future camera-frame pixels as auxiliary training signal. The visual prediction head is dropped at deployment. Tested across four LIBERO simulation suites; the model card notes its cross-paper comparison mixes published protocols.
Robotics API
Gemini Robotics ER 2 →
The embodied-reasoning half of Google DeepMind's new robotics family, and the only part available now - it reasons about physical scenes and plans robot tasks via the Gemini API and AI Studio, while the models that actually drive motors stay in private preview.
Run AI models efficiently
SGLang v0.5.13 →
A high-performance open serving engine for language models. The new version turns on faster 'guess-ahead' decoding by default and trims scheduling overhead for quicker responses.
vLLM v0.23.0 →
The widely-used open engine for serving language models fast and cheaply. The latest release adds smarter memory handling for long conversations and faster GPU execution.
Run models on your computer
LM Studio →
A friendly desktop app to find, download, and chat with open models on your own machine — no command line needed.
Ollama →
Download and run open AI models locally with a single command. The easiest on-ramp to running your own model.
Open WebUI →
A polished, ChatGPT-style web interface for the open models you run yourself.
llama.cpp →
The lean, fast engine that makes big models run on ordinary laptops; powers much of the local-AI ecosystem.
SSD-streaming inference engine
Mference →
Runs DeepSeek V4 Flash on Apple silicon by keeping the shared core, attention and cache resident while streaming each token's routed experts off the SSD. Publishes an unusually honest memory budget: about 3 GB working set, 90 to 98 GB on disk, tested at a 4,000-token context on a 24 GB Mac, with no quality parity test yet.
Safety classifier
Shieldstral 1.0 3B →
Mistral's open-weight multimodal moderation model. You supply the policy as a plain-language yes/no question at inference time rather than retraining for a fixed harm taxonomy, and it returns one calibrated safety score per forward pass. Handles prompts, responses, prompt-response pairs, images and image-plus-text across twelve languages. Apache 2.0, runs on a single 16GB GPU via vLLM, Transformers or llama.cpp; recommended operating context is 32k tokens.
Safety guardrail model
Qwen3Guard →
Alibaba's first open-weights safety-filter model, released under Apache 2.0 in three sizes, covering 119 languages, with a streaming variant that can flag unsafe text token by token as it is generated.
Scientific literature search
AskChem →
A live chemistry search system that retrieves individual claims rather than papers - 2.4 million typed claims from 147,000 papers, each carrying a source identifier and a verbatim quote. Free web interface plus REST, SDK and MCP access so agents can query the same claim store.
Search agent
ABSeeker →
A released 4-billion-parameter web-research agent trained with per-step credit assignment that matches roughly 30-billion-parameter agents on hard fact-finding tasks.
Security coding assistant
OpenAI Codex Security (Daybreak) →
An in-IDE plugin from OpenAI's Daybreak initiative that finds, validates, and fixes software vulnerabilities, plus an open-source remediation program run with Trail of Bits and HackerOne.
Self-hosted AI companion
AIRI →
Self-hosted embodied assistant with a Live2D or VRM character, voice, persistent memory, local inference support, and game and chat integrations. A vertical application rather than a general agent framework.
Serve at scale
vLLM →
The popular open engine for serving AI models fast and efficiently when you need to handle real traffic.
Single-GPU inference engine
NInfer →
A focused inference engine that runs Qwen 3.6 models on one RTX 5090 with a 262,000-token context using an INT8 key-value cache, reporting roughly 188 tokens per second at 250,000 tokens of context. Methodology, seeds and limits are published openly.
Skills library for coding agents
agent-skills →
Addy Osmani's collection of production-grade, reusable skills for AI coding agents, trending near the top of GitHub this week.
Small language model
LFM2.5-2.6B →
Liquid AI's 2.7B tool-calling model with a 128k context, built as 22 short-convolution layers plus 8 grouped-query-attention layers so most token mixing stays local and cache-friendly. Post-trained inside real agent harnesses for tool use, extraction, retrieval and long-context workflows. The model card explicitly recommends against agentic coding and knowledge-heavy tasks, and it always enters a reasoning mode before answering.
Small looped-transformer model
Nanbeige4.2-3B →
An Apache-2.0 4B model that reuses one 22-layer transformer stack twice for 44 layers of depth from a single set of weights, shipping BF16 weights with SGLang, vLLM, llama.cpp, and Ollama paths for local use.
Small open models
Ternary Bonsai models →
A family of 1.7B, 4B and 8B models built for extreme quantization, shipped in the official group-64 two-bit format that mainline llama.cpp reads. Useful if you want to see what 2-bit inference feels like without converting anything yourself.
Speculative decoding draft model
Kimi-K3-DSpark →
Inferact's draft model for Kimi K3. It proposes seven tokens at a time for K3 to verify and accept or discard, and its block-diffusion backbone shares K3's attention-cache layout so no second cache format is needed. This is the component behind the 21-25 tokens per second measured on a sixteen-node GB10 cluster running the full K3 checkpoint.
Speech model
NVIDIA NemotronLabs VoiceChat 11B →
An open-weight end-to-end full-duplex voice model that listens and speaks simultaneously and calls tools mid-conversation, shipped with both offline inference code and a containerised WebSocket streaming deployment. Needs an NVIDIA GPU with at least 80 GB of memory, and uses a single fixed voice.
Speech-to-text
Cohere Transcribe Arabic →
Open-source (Apache 2.0) Arabic speech-recognition model built for dialects and Arabic-English code-switching, with lower word error rate than Whisper Large V3 on the Hugging Face Arabic leaderboard.
Stealth browser / scraping engine
Fortress →
Open-core stealth Chromium with C++-level fingerprint patches that lets browser agents and scrapers pass bot detection (Cloudflare, DataDome, Turnstile); ships 29 pre-built MCP tools for the agentic web.
Terminal coding agent
Kimi CLI →
Moonshot's Apache-2.0 terminal agent for driving Kimi models from the command line for coding and tool use. Open-source software (distinct from the K3 model weights, due July 27).
Text-to-speech model
Inflect-Micro-v2 →
A complete English speech synthesis stack in 9,356,513 parameters, waveform decoder included, producing 24 kHz mono audio locally with no external vocoder or API. One fixed synthetic male voice, no cloning, flatter prosody than large systems - but it runs anywhere.
Tokenizer library
GigaToken →
An open-source native BPE tokenizer optimized with SIMD byte scanning and instruction-level parallelism, best used for fast offline corpus preparation and bulk token counting rather than end-to-end serving speedups.
Train & fine-tune AI models
veRL →
The open RL post-training framework used by most research labs training reasoning models today. Run GRPO, PPO, and related reward-training methods on your own models.
Training dataset
Recursive-Task-Synthesis →
A public set of 37,484 verified long-horizon terminal-agent tasks, each a runnable bundle with instruction, environment, reference solution and hidden verifier, plus a companion set of 327,000 agent trajectories and three fine-tuned Qwen3.5 checkpoints.
The Stack v3 →
Hugging Face's code corpus, now with source text embedded inline rather than behind identifiers. A 15.9 TB deduplicated, PII-redacted training split of roughly 4.9 trillion tokens, plus a 113.7 TB unfiltered bucket for teams that want to build their own mix.
Training framework
MindSpeed-LLM →
Huawei's official large-model training toolkit for Ascend NPUs, covering distributed layouts, checkpoint conversion and supported model families. Worth reading its support table honestly -- DeepSeekV4-Flash is currently marked Prototype, its label for not-fully-validated features.
Vector search
FastPLAID →
A Rust engine for multi-vector search, the indexing layer that makes late-interaction retrieval fast enough to serve in production instead of only benchmarking well.
Video and audio generation
MiniMax-H3 →
Open weights for MiniMax's omni-modal model that generates four to fifteen second video with native stereo audio. The locally deployable base runs at 768p through diffusers or SGLang; the prompt-interpretation and 2K regeneration stages stay behind MiniMax's API, and the licence excludes the US, EU, UK, and South Korea.
Video evaluation toolkit
FilmOps + FilmBench →
Public benchmark assets for judging generated video on professional film craft instead of generic prettiness. FilmOps ships six specialized operators covering shot scale, composition, camera angle, color and tone, character layout and camera movement; the companion FilmBench dataset supplies prompts reverse-engineered from professionally selected clips, most of which require multi-shot continuity. Authors report weaker agreement with human raters on audio and editing than on visual categories.
Video generation
Seedance 2.5 on Dreamina →
ByteDance's newest joint audio-video model, announced 31 July, generating a single take of up to 30 seconds extendable twice, with white-model control, green-screen editing and camera and blocking controls. Note that the 4K output, 50-reference limit and 180-second beta advertised on this page are marked Coming Soon.
Vidu (Vidu S1 Stream Model) →
A working AI video generator with text-to-video, image-to-video, and reference-to-video modes; the new S1 Stream Model targets real-time, interactive, voice-steerable video at up to 42 FPS/540p on consumer GPUs. Free credits to try, paid plans for more.
Video generation API
Gemini Omni Flash →
Google's new video model offering developers programmable conversational editing - generate and revise clips up to ten seconds by describing changes in words.
Video understanding model
Mage-VL →
Microsoft's codec-native multimodal model that reuses a video file's own bit allocation to pick visual tokens, reporting over 75% fewer tokens and up to 3.5x faster inference than uniform frame sampling. Works with H.264, HEVC and DCVC-RT.
Vision-language model
moondream 3.1 (9B-A2B) →
An open-weight vision-language model with 9B total but only 2B active parameters, offering native object detection, pointing, captioning, and segmentation at roughly the speed of a 2B dense model.
Voice agent framework
LiveKit Agents →
Production framework for building realtime voice agents, with interchangeable speech-to-text, LLM, text-to-speech, and realtime components plus semantic turn detection. This is the plumbing layer around a voice model rather than a duplex model itself, and it trended on GitHub today.
Voice agents
speech-to-speech →
Hugging Face's modular local voice-agent pipeline - voice detection, speech recognition, a language model and text-to-speech chained together, with an OpenAI Realtime-compatible websocket so existing clients can point at it.
Voice pipeline
Hugging Face speech-to-speech →
Local voice-activity detection to speech recognition to language model to text-to-speech pipeline, threaded through queues and exposed as an OpenAI Realtime-compatible server so existing clients can point at it unchanged.
Weather forecasting API
WeatherNext models on Google Cloud →
Google DeepMind's AI weather forecasts, available as a developer API and as raw forecast data in Earth Engine, BigQuery and Vertex AI, with ensemble scenarios out to 15 days.
Web data API
Firecrawl →
A hosted API that crawls, scrapes and structures web pages into clean text for agents and retrieval pipelines, handling the JavaScript rendering and rate limiting you would otherwise build yourself.
Web publishing control
Cloudflare AI Bot Controls →
A free Cloudflare setting, live since July 1 2026, that lets any site separately allow or block three kinds of AI crawler: search indexers, live AI assistants, and model-training scrapers.
Website bot policy
Cloudflare AI crawler controls →
Free-tier controls that split AI crawler traffic into Search, Agent and Training, each set independently to allow, block site-wide, or block only on ad-bearing pages. These are edge blocks on classified traffic, not robots.txt requests. New domains change default on September 15, 2026.
Writing-style leaderboard
The Slop Index →
An open leaderboard scoring how much like generic AI prose a model writes, combining blind pairwise crowd votes with mechanical style measures against a pre-2022 human reference corpus. Methodology and generations are public; treat the rankings as a prototype, since the project's own published counts do not reconcile.