long-context
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.
Linear Attention Lesson
Linear attention rewrites the transformer's attention math so its cost grows in a straight line with sequence length instead of exploding as the square, letting models handle very long inputs cheaply, at the price of a compressed memory of the past.
LongStraw Makes Million-Token RL Training Possible on 8 GPUs - But Its Code Doesn't Run Yet News
A new systems paper called LongStraw shows reinforcement-learning post-training can execute on prompts beyond 2 million tokens on a fixed 8-GPU budget by scoring the shared prompt once without gradients and backpropagating only through the short generated response, though the authors call this proof of execution capacity rather than full training correctness, and the public code is not yet runnable.
State Space Models and Mamba Lesson
State space models are a transformer alternative that processes a sequence by carrying a fixed-size running summary forward one step at a time, giving them linear cost with length and constant memory per token -- which is why models like Mamba can handle very long inputs cheaply.
A new attention method learns what to ignore, and reads 64x more text News
Tencent's Hunyuan team introduced HiLS, a sparse-attention method that learns end-to-end which parts of a long document to focus on, matching full attention while handling context 64 times longer than it was trained on.
New methods page an AI's memory in and out like an operating system News
Two new techniques treat a language model's long-context memory like an operating system's memory hierarchy - keeping coarse summaries on the GPU and paging compressed detail out to the CPU - with one, SeKV, cutting GPU memory use by 53% at 128,000 tokens while avoiding the errors that plague simpler methods.
Z.ai's GLM-5.2 claims the open-weight coding crown with a usable 1M-token context News
Z.ai released GLM-5.2, an agentic coding model with a reliable one-million-token context and top open-source scores on long-horizon software benchmarks, with an MIT-licensed weight release promised within weeks.
Sparse Attention Lesson
Sparse attention lets a transformer skip most of the pairwise comparisons between tokens, so instead of every token attending to every other token, each one attends to a chosen subset -- which is what makes million-token context windows affordable.
The KV cache: why AI gets slower and hungrier the longer it talks Lesson
The hidden notebook that lets a model avoid re-reading every previous word - and the single biggest reason long context is expensive.
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.
What is a context window? Lesson
A model's context window is how much text it can hold in mind at once — its working memory. Bigger is useful, but a long window isn't the same as a good memory. Here's how it works and where it breaks.
vLLM DeepSeek-V4 support Tool
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.
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.
Qwen3.8-Max Tool
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.
Qwen3.6-35B-A3B Tool
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.
Qwen3-Next-80B-A3B-Instruct Tool
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.
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.
NInfer Tool
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.
MiniMax-M3 Tool
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.
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.
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.
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.
Gemini 3.6 Flash and 3.5 Flash-Lite Tool
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.
GLM-5.2 Tool
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.
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 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.