open-weights
Z.ai changed only the post-training, and the model learned to find exploits News
Z.ai released GLM-5.3 on August 14 using the same base model as GLM-5.2, with every gain coming from post-training, and the largest jump was in finding and exploiting software vulnerabilities.
Qwen3.8-27B shares its predecessor's bones, but not its contract News
Alibaba's Qwen3.8-27B shipped with the same coarse architecture as Qwen3.6-27B, prompting accusations it was a relabel with knowledge stripped out, but the published comparison shows knowledge scores flat or slightly up.
MiniMax released a five-minute song model with a catch in the licence News
MiniMax published the weights for Music 3, a model that generates complete five-minute songs with vocals in 32 kHz stereo, under a licence that permits commercial use but requires on-screen credit and written permission above $20 million in revenue.
DeepSeek starts charging rush-hour prices on August 17 News
DeepSeek is replacing flat API pricing with peak and off-peak rates on August 17, and the steepest change hits cached input on its Pro model, which goes up twelvefold during Beijing business hours.
Chinese models passed American ones in OpenRouter traffic in June News
OpenRouter's own analysis dates the crossover where Chinese models overtook American ones in token share to early June 2026, driven by DeepSeek V4 Flash taking 70 percent of DeepSeek's agentic traffic.
DeepSeek's new open model is 1.6 trillion parameters and runs 49 billion of them per token News
DeepSeek published DeepSeek-V4-Pro on Hugging Face with 1.6 trillion total parameters, 49 billion activated per token, and a one-million-token context window, making it the largest openly downloadable model of the current frontier wave.
NVIDIA built a 30B model for the boring half of agent work News
NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model that activates only 3 billion parameters per token and is aimed at the tool calls and validation steps that consume most of a long-running agent's time.
Macaron froze a 744B base and bolted four specialists on top News
Mind Lab released open weights for Macaron-V1, a continual-learning system that never touches its base model and instead composes small specialist adapters on top, picking exactly one per user turn.
LTX-2.5 ships open weights and a chart that races its own hardware News
Lightricks released LTX-2.5 with downloadable weights, free commercial use under $10M in revenue, and a rebuilt generation stack whose new diffusion decoder spends rendering compute where the scene is complex.
Mistral patented letting the model write the tool call as code News
Mistral AI holds a granted US patent, "Code implemented tool calls," covering an agent architecture in which a model writes a code block wrapping tool calls, a server runs it in a sandbox, pauses at each external call, and resumes with the result substituted in.
Meta ships a 30B agent model that fits on one gaming GPU News
Meta released Muse Glimmer, a 30-billion-parameter open-weight agent model under the permissive Apache 2.0 license that shrinks to under 20 GB once quantized, small enough to run on a single 24 GB consumer graphics card.
The open video model tops out at fifteen seconds, not twenty-six News
MiniMax's H3 weights have drawn 2,900 stars in four days, but the company's own repository caps a single generation at fifteen seconds and says the hosted component it left out is critical to output quality.
The harness, not the model, moved DeepSeek's score by twenty tasks News
Independent benchmarker Harrison 'sentdex' Kinsley re-ran DeepSeek V4 Flash 0731 on the same 89-task terminal benchmark under a different agent harness and watched it go from 44 solved to 64 solved, without changing the model.
Google put the hurricane model's weights on a public bucket News
DeepMind has open-sourced WeatherNext 2 with pretrained weights, including the exact cyclone checkpoint the National Hurricane Center used operationally through the 2025 Atlantic season, plus a one-degree variant sized for a single consumer GPU.
Sixteen AI-designed viruses worked, and one borrowed a part from a cousin News
Arc Institute researchers used a genome language model to design bacteriophages from scratch, synthesized the DNA, and got 16 working viruses out of 285 tested -- one of which swapped in a structural protein from a distantly related phage.
The White House's Open-Weight Carve-Out Is a Private Briefing, Not a Published Rule News
Reporting says the White House finished an AI framework that covers only closed frontier models and will not publish it, but the only public legal instrument is June's Executive Order 14409, which contains no definition of open-weight, no US-origin condition and no mandatory testing regime to be exempt from.
The Full 2.8-Trillion-Parameter Kimi K3 Now Runs on Sixteen Desktop Boxes News
An operator has the complete Kimi K3 checkpoint running across sixteen GB10 mini-workstations wired through a single 400G switch, producing roughly 21 to 25 tokens per second for one user, on hardware with a verifiable floor around $57,200.
The 'Ternary' 20B Model Everyone Downloaded Today Ships as a Two-Bit Package News
DeepGrove's Maple-Preview is being shared as a 20-billion-parameter ternary model that fits in about 5 gigabytes, but its native release is a 40GB BF16 repository and the small footprint belongs to a separate two-bit MLX build with no published ternary training recipe.
Mistral Shipped an Open-Weight Safety Judge That Takes Its Policy as a Question News
Mistral released Shieldstral 1.0 3B, an Apache-2.0 multimodal moderation model that reads a plain-language yes/no policy question at inference time instead of a fixed harm taxonomy baked into its weights, and runs on a single 16GB GPU.
Liquid Shipped a 2.6B Tool-Calling Model and Told You Not to Code With It News
Liquid AI released LFM2.5-2.6B, a small model whose stack is mostly cheap convolution layers with attention rationed to eight of thirty, built for tool use and long context - and whose own model card says not to use it for agentic coding or knowledge-heavy work.
Qwen3.8-Max Shipped as a Paid API, Not as Open Weights News
Alibaba put Qwen3.8-Max live as a hosted API at $2 per million input tokens and $6 per million output tokens, a fifth cheaper than the model it replaces, while the open weights it promised for Max and a 27B sibling have not shipped.
NVIDIA's Open Full-Duplex Voice Model Wants an 80GB GPU News
NVIDIA released an 11-billion-parameter speech model that listens and speaks at the same time and calls tools mid-conversation, and its own documentation requires a GPU with at least 80 GB of memory and lists more than a dozen failure modes.
MiniMax Shipped H3's Weights and Kept the Best Part Hosted News
MiniMax released the weights for its H3 video-and-audio generation model, and its own model card says the input-processing stage that is critical to output quality is not included in the release and the 2K output stage is not open-sourced at all.
llama.cpp ships the fix that lets DeepSeek V4 Flash call tools mid-thought News
DeepSeek's new open-weight model shipped without a standard chat template, and local agents kept dying when it emitted a tool call inside its reasoning; llama.cpp release b10217 and KoboldCpp v1.118 both landed fixes on 1 August.
China did not give away free models. It built a governance body. News
Reports that China offered free AI models to the Global South at a Geneva summit describe a discussion session; the concrete instrument came nine days later in Shanghai with the founding of an intergovernmental AI cooperation organisation.
Metis puts an agent's memory inside the model instead of in a database beside it News
A team spanning several Chinese universities released the first prototype of what it calls a memory foundation model - a backbone carrying a memory state that updates on every interaction through a plain forward pass, with no gradients and no external retrieval system.
DeepSeek re-trained V4 Flash without touching the architecture and its coding-agent score went from 7 to 54 News
DeepSeek published new MIT-licensed weights for V4 Flash on July 31 that change only the post-training, lifting the model's score on a real-world software-engineering agent test from 7.3 to 54.4 out of 100.
An open 35B model trained to evolve its own machine-learning code nearly doubled its base model's medal rate News
Frontis-MA1, released with full weights and stack, raises its base model's medal average on a machine-learning engineering benchmark from 39.4% to 60.6%, and to 71.2% with a stronger search - all within a 12-hour budget on a single consumer GPU capped at 12GB.
Thinking Machines ships Inkling-Small's open weights - all 532 gigabytes of them News
Thinking Machines has published the full weights for Inkling-Small, a 276-billion-parameter sparse model that activates only 12 billion parameters per token and accepts text, images and audio, under an Apache 2.0 licence with a separate use policy attached.
LG shipped a 750-billion-parameter model and quietly dropped its restrictive licence News
LG AI Research released K-EXAONE 2.0, a 750-billion-parameter sparse model with 37 billion active, under Apache 2.0 - a break from the custom EXAONE licence that governed its previous releases.
Researchers built a model whose dangerous knowledge can be switched off like a module News
A method called GRAM routes risky training data into small auxiliary modules that can be turned on or off after training, so one model can approximate several models each trained without a different category of dangerous data, tested from 50 million to 5 billion parameters.
A $500 fine-tune of a 9B open model beat all five frontier models it was tested against News
A consultancy reinforcement-trained a 9-billion-parameter open model on a simulated product-catalog review workflow for about $500 of GPU time, and it outscored the best of five frontier configurations while costing $0.50 per thousand listings against $34.
Microsoft lets the video codec pick which pixels the model sees News
Microsoft's Mage-VL reuses a video file's own compression decisions to choose which image patches a vision model processes, cutting visual tokens by over 75% and reporting up to a 3.5x speedup over uniform frame sampling.
Kimi K3 topped a fullstack coding board at maximum effort News
Moonshot's open-weight Kimi K3, served at its highest reasoning setting, took first place on Code Arena's July 23 WebDev snapshot over Claude Fable 5 and GPT-5.6 Sol, though the live board has since moved it to second.
DeepSeek V4 Flash hits 32 tokens a second on one desktop News
A published benchmark shows DeepSeek's 284-billion-parameter V4 Flash generating 32 tokens per second entirely on one AMD Strix Halo machine, using aggressive quantization, speculative decoding and reduced expert routing.
Chinese open models passed US models in OpenRouter token share News
OpenRouter's own usage data shows Chinese models overtaking US models in token volume in early June, with DeepSeek roughly doubling its share to 18% - driven by token-hungry agent workloads routing to the cheapest capable endpoint.
NVIDIA launches an open AI security alliance with 41 partners, and OpenAI is not on the list News
NVIDIA announced the Open Secure AI Alliance with 41 inaugural partners including Microsoft, the Linux Foundation, Hugging Face and CrowdStrike, built around the claim that closed APIs blocked forensic work during the Hugging Face breach while an open model did it.
Moonshot releases Kimi K3: a 2.8-trillion-parameter open-weight model, 1.56 terabytes on disk News
Moonshot AI published the full downloadable weights for Kimi K3, a 2.8-trillion-parameter model that uses only 104 billion parameters per token, handles text and images, and reads just over a million tokens of context.
Kimi K3 is downloadable, but the floor to run it is eight datacenter GPUs News
Kimi K3's 1.56-terabyte checkpoint needs a single eight-GPU B300 or MI355X node as its practical minimum, and no version of llama.cpp can load it today, so open weights currently mean operator-scale rather than local.
Anthropic says it never asked to ban open-weight models, and names what it does want instead News
Anthropic published its position on open weights, rejecting a categorical ban while backing three specific restrictions: chip export controls, action against industrial-scale distillation, and mandatory pre-release safety testing for sufficiently capable models, open or closed.
llama.cpp Merges MiniMax M3's Sparse Attention, Because Running It Dense Gives Wrong Answers News
Support for MiniMax M3's block-sparse attention landed in llama.cpp today, and the contributor is explicit that it is not a speed optimisation: the model was trained sparse, so running it dense produces degraded output.
Upstage Ships Solar Open 2, a 250-Billion-Parameter Model That Runs on Two GPUs News
Korea's Upstage released open weights for a 250-billion-parameter agent model that activates only 15 billion parameters per token and fits on two NVIDIA H200s when quantized, under a custom license with attribution requirements.
Lobbying Filings Show Anthropic Named Distillation and Export Controls. OpenAI's Did Not. News
After the New York Times reported that both labs privately pressed Washington over Chinese open-weight models, their own second-quarter lobbying disclosures tell sharply different stories about what each one admits to working on.
Kimi K3's Open Weights Are Still a Countdown, Not a Download News
On the eve of its promised release, Moonshot AI's Kimi K3 page on Hugging Face is a timer with no weights, no license file, and no technical report behind it.
Baidu's OCR Model Reads Dozens of Pages in One Pass Without Its Memory Growing News
Unlimited OCR replaces the decoder's attention layers with a design that holds the key-value cache at a constant size, so transcription speed does not degrade as the document gets longer.
A Popular Jailbroken Gemma 4 Shipped With 54 Attention Tensors Missing News
The publisher of a widely downloaded guardrail-stripped Gemma 4 admits its earlier version silently deleted 54 shared attention tensors, producing hallucinations that users had no way to distinguish from ordinary model weakness.
llama.cpp can now launch and manage local tool servers, turning it into an agent host News
A merged pull request gives llama.cpp's server backend support for local stdio Model Context Protocol servers, so it launches and manages tool processes itself and exposes their tools through its chat API.
The open-weights letter doubled to 50 signatories - Google and OpenAI signed, Anthropic did not News
The industry letter urging Washington not to restrict open-weight AI doubled from 25 signatories to 50 within a day, adding Google and OpenAI; Anthropic is not on the list.
DeepSeek paused its funding round after a leaked meeting transcript went viral News
Bloomberg reports DeepSeek told prospective backers it would not sign expected agreements for now, a suspension its sources attribute in part to viral posts about a leaked investor-meeting transcript attributed to founder Liang Wenfeng.
A complete text-to-speech system now fits in 9.4 million parameters News
Inflect-Micro-v2 packs an entire English speech synthesis stack, including the waveform decoder, into 9,356,513 parameters that run locally with no external vocoder or API.
VideoChat3 Halves Video-Model Latency by Compressing Space and Time First News
VideoChat3, an open 4-billion-parameter video model, compresses frames across space and time before the language model, roughly halving latency versus a comparable model.
The Director of the U.S. AI-Evaluation Agency Is Leaving After Three Months News
CAISI Director Chris Fall is leaving after about three months, with NIST Director Arvind Raman becoming acting head, days after the agency published a detailed assessment of a Chinese open-weight model.
Tencent's Open Robot Model Plans by Imagining the Scene It Wants to Create News
Tencent released RxBrain, an open ~6.2B robot model that interleaves text reasoning with generated goal images, betting that a robot needs an explicit picture of the world it is trying to build.
Safety Guardrails Blocked a Security Team's Own Incident Analysis News
Hugging Face disclosed that commercial AI safety filters blocked its analysis of real attack code during an incident, so it ran the forensics on a self-hosted open-weight model instead.
Axios: U.S. Officials Revive an Effort to Discourage Chinese Open-Weight AI News
Axios reports that internal U.S. efforts to restrict Chinese open-weight AI models have revived after Kimi's rise, but no ban, rule, or executive order has been announced.
Boogu-Image-0.1: a fully open image model that claims to close in on closed systems for about $400K News
Boogu-Image-0.1 is a fully open-source unified image generation and editing model family whose researchers say a base model reaching near-frontier quality cost roughly $400,000 to train, arguing the closed-open gap is closing through data and pipeline quality rather than raw compute scale.
Thinking Machines releases Inkling, now the top-ranked US open-weights model News
Thinking Machines Lab released Inkling, a 975-billion-parameter open-weights model under Apache 2.0 that Artificial Analysis ranks as the strongest open-weights model from any US lab, scoring 41 on its Intelligence Index.
A leading open-model researcher says US open weights may have six months left News
Nathan Lambert of the Allen Institute argues in a widely-read essay that a coming White House executive order could ban or indefinitely delay any open-weights model above roughly GPT-5.5 capability, and that the industry's distillation debate is regulatory capture.
What Ring-2.6-1T's model card actually says News
Ant Group's openly downloadable trillion-parameter model is real and MIT-licensed, but its benchmark claims are vendor-supplied and measured against a previous generation of rivals -- not the current frontier.
Bonsai puts a 27B model on your phone, and shows what breaks News
PrismML shrank a 27-billion-parameter model from 54 gigabytes to 3.9 by storing each weight as a single bit, and published the benchmark table showing tool calling and vision falling off a cliff.
George Hotz: I love LLMs, I hate hype -- and the labs won't capture the value they create News
Hacker and comma.ai founder George Hotz argued that AI is real and useful but that frontier labs will fail to capture the value they create, because AI is a continuation of the general computer revolution rather than the labs' unique moat.
Meituan open-sources LongCat-2.0, a trillion-parameter model it says was trained end-to-end on Chinese chips News
Meituan released LongCat-2.0, a 1.6-trillion-parameter open-weight (MIT) model that ran anonymously as 'Owl Alpha' for two months and was, the company says, both trained and served entirely on domestic Chinese AI ASICs with no Nvidia GPUs.
Cohere open-sources an Arabic speech model that beats Whisper and handles dialects and code-switching News
Cohere released an open-source (Apache 2.0) Arabic speech-recognition model that it says is the most accurate open Arabic transcriber to date, cutting word error rate well below OpenAI's Whisper Large V3 and handling dialects and Arabic-English code-switching, not just formal Modern Standard Arabic.
Alibaba's Qwen3Guard flags unsafe AI output token-by-token as it's being generated News
Alibaba released Qwen3Guard, its first open safety-guardrail model, whose streaming variant classifies an AI response for safety as each token is generated rather than after the fact -- and adds a 'Controversial' tier between Safe and Unsafe that apps can tune stricter or looser.
Tencent open-sources Hy3, a lean mixture-of-experts model that punches above its weight News
Tencent released Hy3 under the permissive Apache 2.0 license: a mixture-of-experts model with 295 billion total but only 21 billion active parameters and a 256K context window, which the company says competes with models five times its size.
Google's Gemma 4 is a small open multimodal family that skips the image encoder News
Google released Gemma 4, an open-weight model family from 2.3 to 31 billion parameters that natively handles vision and audio, including a 12-billion-parameter variant that ingests raw image and audio patches with no separate encoder.
Chinese open models now handle a third of US enterprise AI traffic News
US companies now route more than 30% of their AI tokens through Chinese open-weight models like DeepSeek and GLM-5.2 every week since February, peaking near 46%, up from an 11% average the year before, according to CNBC's analysis of OpenRouter data.
A $4-per-million open model is coming for the frontier's 90% margin News
GLM-5.2, an open-weights model priced at under a fifth of Opus, scores as the top open model and 4th overall - and a widely-shared essay argues it is the first real threat to frontier labs' ~90% inference margins.
Together AI raises $800M as open-source AI inference crosses $1B News
Together AI raised an $800 million round on July 1 at an $8.3 billion valuation, more than doubling in about 16 months, as its annual bookings crossed $1.15 billion and open-source model usage tripled in a year.
This method compiles plain English into a tiny model that rivals a 32B giant News
A new paper, Program-as-Weights, uses a 4B compiler to turn natural-language specs into small neural adapters that let a 0.6B model match a 32B model on fuzzy tasks, running at about 30 tokens per second on a laptop.
DeepSeek's new open models give everyone a million-word memory by default News
DeepSeek previewed two free-to-download V4 models that can read a million tokens at once, no longer as a premium add-on but as the standard setting.
Distillation: how a small AI learns from a big one Lesson
Distillation trains a smaller, cheaper model to imitate a larger, smarter one, the idea behind both efficient deployment and the 'copying' accusations now driving AI geopolitics.
Are closed AI models overpriced luxury goods? News
An essay argues open-weight models now undercut the big closed AIs by huge margins, and that 'China fears' are being used to protect those prices.
A language model that doesn't write left to right News
iLLaDA is an 8-billion-parameter model that generates text by refining a blurry whole rather than one word at a time, and it's catching up to the mainstream.
moondream 3.1 (9B-A2B) Tool
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.
Xiaomi MiMo-V2.5-DFlash Tool
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.
WorldDiT Tool
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.
WeatherNext 2 Tool
Google DeepMind's ensemble weather and cyclone forecasting models, released with code, pretrained weights and runnable notebooks. Includes the checkpoint used operationally by the National Hurricane Center in 2025 and a one-degree Mini variant sized for a single GPU.
VideoChat3-4B Tool
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.
Unsloth Qwen3.8-27B GGUF Tool
Quantized builds of Alibaba's newest 27B open-weight model, published within minutes of the release, in a range of sizes that fit on a single consumer graphics card.
Unsloth Kimi-K3-GGUF Tool
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.
Unsloth DeepSeek-V4-Flash-0731 GGUF Tool
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.
Unlimited OCR Tool
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.
Ternary Bonsai models Tool
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.
Solar Open 2 Tool
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.
Shieldstral 1.0 3B Tool
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.
RxBrain (Hy-Embodied-RxBrain-1.0) Tool
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.
Ring-2.6-1T Tool
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.
Qwen3Guard Tool
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.
Qwen3.6 (open weights) Tool
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.
Qwen-Image-2.0-Pro Tool
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.
Poolside Laguna S 2.1 Tool
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.
OS-Shepherd-9B Tool
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.
Neuronpedia J-lens demo Tool
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.
Nemotron-3-Puzzle-75B Tool
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.
Nanbeige4.2-3B Tool
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.
NVIDIA NemotronLabs VoiceChat 11B Tool
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.
NVIDIA Nemotron 3.5 Lightning Tool
A 30B mixture-of-experts model with only 3B parameters active per token, trained for the high-volume half of agent work: tool calls, result validation and subagent delegation. NVFP4 and BF16 checkpoints, with weights, training data and recipes released under a permissive licence.
NVIDIA Nemotron 3 Embed 8B Tool
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.
Muse Glimmer 30B Tool
Meta's 30-billion-parameter open-weight agent model under Apache 2.0, built for always-on local workflows with text and image input, tool use, a context window past 131,000 tokens, and a speculative decoder that drafts sixteen words at a time. Full weights, quantized builds, and the drafter are all in the release.
MiniMax-H3 Tool
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.
MiniMax Music 3 Tool
Open-weight model that generates complete five-minute songs with vocals in 32 kHz stereo from lyrics plus a structured style description. Runs via SGLang-Omni, Diffusers, or ComfyUI. Commercial use allowed with on-screen attribution; written permission required above $20M revenue.
MiniCPM5-1B Tool
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.
Maple-Preview (2-bit MLX build) Tool
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.
Macaron-V1 Tool
Open weights for a model family that freezes its base and composes specialist LoRA adapters on top, picking one per user turn. The 744B Venti flagship carries chat, agent, coding and generative-UI specialists; the 50B Tall variant runs the same design on local hardware.
LongCat-2.0 Tool
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.
Ling-3.0-flash Tool
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.
LateOn Tool
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.
LTX-2.5 Tool
Lightricks' open-weight video foundation model, shipped 11 August with a new diffusion decoder, native multishot generation, 4K HDR and automatic clip-length prediction. Runs locally on 16GB of VRAM, or through a per-second API. Free for commercial use below $10M annual revenue.
LLaDA / iLLaDA Tool
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.
LFM2.5-2.6B Tool
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.
Kimi K2.7 Code Tool
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.
Inkling-Small GGUF Tool
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.
Inkling Tool
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.
Inflect-Micro-v2 Tool
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.
Google Gemma (open weights) Tool
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.
Gemma 4 26B A4B Tool
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.
Gemma 4 Tool
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.
Frontis-MA1-35B Tool
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.
Evo 2 Tool
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.
DenseOn Tool
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.
DeepSeek-V4-Pro Tool
A downloadable 1.6-trillion-parameter mixture-of-experts model that activates 49 billion parameters per token, with a one-million-token context window under an MIT license. Serious server hardware required, but the weights are yours.
DeepSeek-V4-Flash-0731 GGUF (Unsloth) Tool
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.
DeepSeek-V4-Flash Tool
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.
DeepSeek-V4 (Pro & Flash) Tool
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.
DeepSeek V4 Pro (API) Tool
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.
DeepSeek V4 Tool
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.
Cosmos3-Edge Tool
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.
Cohere Transcribe Arabic Tool
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.
Bonsai 27B (GGUF) Tool
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.
ABSeeker Tool
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.