Long Context

Long Context refers to models and techniques that process very large inputs — many thousands to millions of tokens — in a single pass, letting a system reason over whole documents, codebases, or conversations at once.

A large window is not free capability: attention can weaken toward the middle of long inputs (the lost-in-the-middle effect), latency and cost rise with length, and more tokens do not guarantee better answers. Bigger is not automatically a substitute for retrieval.

For builders the practical question is long context versus RAG: loading everything is simpler but slower and pricier, while retrieval is leaner but can miss. Most production systems combine them, using Context Engineering to place only the most relevant material into even a large window.

Topic: LLMs Related: Context Engineering RAG

Recent Updates

• 2026-07-16: Adaptive KV Cache Filtering for Long-Context Inference (cs.CL updates on arXiv.org) · arxiv.org
• 2026-07-16: Thinking Machines Lab releases Inkling, an open-weights multimodal model family (Latent.Space) · latent.space — Thinking Machines Lab · Latent.Space · Hugging Face · vLLM · SGLang · Modal · Baseten · Databricks Mira Murati Soumith Chintala John Schulman Lilian Weng · Inkling · Inkling-Small
• 2026-07-16: An Essay on Why AI Memory Should Be a Separate Layer (量子位) · qbitai.com — QbitAI Shadoweave · OpenAI Sutton’s Oak Lab H.M. Solomon Shereshevsky Ted Chiang Wenle Sutton
• 2026-07-16: NVIDIA releases Nemotron 3 Embed for retrieval-focused RAG and agent workflows (Hugging Face - Blog) · huggingface.co — NVIDIA · Hugging Face · NVIDIA NeMo · vLLM NVIDIA NIM Nemotron 3 Embed Nemotron-3-Embed-8B-BF16 Nemotron-3-Embed-1B-BF16 llama-nemotron-embed-vl-1b-v2 · Nemotron 3 Ultra
• 2026-07-16: Kimi Code Adds K3 Configuration and Switching Guidance (Hacker News) · kimi.com — Kimi K3 · Kimi-K2.7-Code
• 2026-07-16: Diagnosing and Mitigating Context Rot in Long-horizon Search (@askalphaxiv) — Shijie Xia Yikun Wang Zhen Huang Pengfei Liu
• 2026-07-16: Diagnosing and mitigating context rot in long-horizon search (@askalphaxiv) · x.com — alphaXiv Fudan University Shanghai Jiao Tong University Shijie Xia Yikun Wang Zhen Huang Pengfei Liu · GPT-OSS 120B · Qwen3.5-397B-A17B · GLM-4.7 GLM-5.0 · MiniMax-M2.5
• 2026-07-16: Kimi K3 launches as a 2.8T-parameter open model with 1M-token context (Hacker News) · kimi.com — Kimi · NVIDIA · Kimi K3 · Claude Fable 5 · GPT-5.6 Sol · Kimi K2 · Opus 4.8 · GPT 5.5
• 2026-07-17: Moonshot AI launches Kimi K3 as a 2.8T open-weights model with 1M context (Latent.Space) · latent.space — Moonshot AI · z.ai · Latent Space · AINews · Arena · Artificial Analysis SimonW Jianlin_S Yulun_Du scaling01 eliebakouch kimmonismus nrehiew_ · Kimi K3 · Claude Fable 5 · GPT-5.6 Sol · Opus 4.8 · kimi-k2.6 · GPT 5.5
• 2026-07-17: Local LLM deployment for SOC: how much GPU is really needed? (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — R-Vision Sergey Ivanov Qwen3.5-122B-A10B-GPTQ
• 2026-07-17: NVIDIA paper argues longer context helps embodied AI policies (DAIR.AI) — NVIDIA RoboTTT
• 2026-07-18: Shishi Technology launches Vectron, a token optimization platform for domestic chip stacks (量子位) · qbitai.com — 是石科技 拓元 Vectron 量子位 · QbitAI · WAIC 2026 昇腾 昆仑芯 天数智芯 太初 瀚博半导体 摩尔线程 · 沐曦 燧原 · 清华大学 · 北京大学 北京航空航天大学 闫博文 允中
• 2026-07-18: Hybrid SWA production results for long-context inference (alphaXiv) · x.com — Xiaomi · alphaXiv · Mimo v2.5
• 2026-07-19: Last Week in AI: China, Compression, and the Open-Model Race (TheSequence) · thesequence.substack.com — TheSequence · Thinking Machines Lab · Moonshot AI · PrismML · OpenAI · NVIDIA · Google Shanghai World AI Conference Xi Jinping · Inkling · Kimi K3 · Bonsai 27B GPT-Red · GPT-5.1 · GPT-5.6
• 2026-07-20: Baidu’s Unlimited-OCR is claimed to read full documents in one pass (Yann LeCun) · x.com — Baidu · Hugging Face · GitHub · Amazon · Google · Azure Yann LeCun Vaibhav Sisinty · Unlimited-OCR
• 2026-07-21: SelKV: Training-Free KV Cache Compression with Merge-or-Drop and Attention Compensation (cs.AI updates on arXiv.org) · arxiv.org
• 2026-07-21: Shapley Context Pruning: Game-Theoretic Reranking and Pruning for RAG Contexts (cs.AI updates on arXiv.org) · arxiv.org
• 2026-07-21: Laguna S 2.1 targets long-horizon agentic coding with 1M-token context (Hacker News) · poolside.ai — Poolside · Datacurve · Tencent · Laguna S 2.1 Laguna M.1 XS.2 DeepSWE Terminal-Bench 2.1 SWE-bench Multilingual · SWE-bench Pro SWE Atlas Toolathlon Verified Tencent Hy3 · Inkling · Nemotron 3 Ultra · DeepSeek-V4-Pro-Max · Kimi K3 · Qwen 3.7 Max · Muse Spark 1.1 · Claude Fable 5
• 2026-07-22: Fusion Embedding proposes one shared embedding space for text, image, video, and audio (cs.CL updates on arXiv.org) · arxiv.org — arXiv Abdul Basit Tonmoy fusion-embedding-1 fusion-embedding-2
• 2026-07-24: TRSP adds a parameter-free side path to reduce representation collapse in long-context Transformers (cs.AI updates on arXiv.org) · github.com — Meta Yiheng Tao Kaiwen Cheng Yao Lu Chang Liu Jie Chen · LLaMA · Llama-3.2-1B
• 2026-07-24: Domyn-Small: 10B open-weight reasoning model with 32K native context and YaRN extension to 128K (cs.CL updates on arXiv.org) · arxiv.org — Domyn Domyn-Small Qwen3.5-9B Olmo-3-7B-Think Nemotron-Nano-8B Ministral-3-8B
• 2026-07-24: REFACT proposes adaptive fact restatement for grounded reasoning (cs.CL updates on arXiv.org) · github.com — arXiv · GitHub
• 2026-07-24: Opus 5 tops the Artificial Analysis Intelligence Leaderboard (Hacker News) · artificialanalysis.ai — Artificial Analysis · Claude Opus 5 · Claude Fable 5 · GPT-5.6 Sol Mercury 2 HyperNova 60B 2605 Granite 4.0 H Small Gemma 3n E4B Instruct Nova Micro Sarvam 30B · Gemini 2.5 Flash Lite · Command A+ · Gemini 2.5 Flash · GLM-5.2 · MiniMax M3 · DeepSeek V4 Pro
• 2026-07-25: A debate on whether agents are already useful or still just getting started (量子位) · qbitai.com — QbitAI · Alibaba Kujing Technology Miaopai CCF · Tsinghua University Nanjing University Shanghai Jiao Tong University Tianjin University Shandong University Shanghai AI Laboratory Jin Lei Gao Yang Hao Jianyie Han Zhongyi Wen Ying Zhou Hao Zhang Hangfan Du Yanlong
• 2026-07-27: Kimi-K3: Open-Source Frontier Model with Native Agentic and Coding Capabilities (Hacker News) · huggingface.co — Moonshot AI · Hugging Face · Kimi K3

FAQ

What is Long Context?

Long Context refers to model and system designs that process large input windows while preserving useful attention and answer quality. GROUNDING tracks context limits, compression, memory, and retrieval-vs-context tradeoffs.

Which topic does Long Context belong to?

On the GROUNDING radar, Long Context is grouped under the LLMs topic.

Which concepts are related to Long Context?

Related concepts tracked by the radar include Context Engineering, RAG.