Tech & AI

The AI Models with the Longest Context Windows

The AI language models that can take in the most text at once, one per lab, ranked by the longest context window in each lab's official documentation as of October 18, 2026. Three models now accept 10 million tokens — Alibaba's Qwen-Long, Meta's Llama 4 Scout and Pokee AI's Pokee-Isaac 28B — while most of the industry, from OpenAI and Google to Anthropic and DeepSeek, has settled at about 1 million.

Video thumbnail about the 10-million-token context window of Meta's Llama 4 Scout

Pictured: video thumbnail from The AI Daily Brief on Meta's Llama 4 Scout, tied for #1 with a 10-million-token context window.

Rank Model Context window (tokens) Maker Country Weights Released Max output (tokens)
1Qwen-Long10MAlibaba (Qwen)ChinaClosedMay 20248,192
1Llama 4 Scout10MMetaUnited StatesOpenApr 2025—
1Pokee-Isaac 28B (preview)10MPokee AIUnited StatesClosedAug 202660,000
4MiniMax-Text-014MMiniMaxChinaOpenJan 2025—
5Amazon Nova 2 Lite1MAmazonUnited StatesClosedDec 202565,536
5Ling-2.5-1T1MAnt GroupChinaOpenFeb 2026—
5DeepSeek-V4-Pro1MDeepSeekChinaOpenApr 2026384K
5Grok 4.31MxAIUnited StatesClosedApr 2026—
5Nemotron 3 Ultra1MNvidiaUnited StatesOpenJun 2026—
5Kimi K31.05MMoonshot AIChinaOpenJul 2026—
5Solar Open 21MUpstageSouth KoreaOpenJul 2026—
5GLM-5.31MZ.ai (Zhipu)ChinaOpenAug 2026128K
5Hy4 preview1MTencentChinaOpenAug 2026—
5Claude Fable 5.11MAnthropicUnited StatesClosedSep 2026128K
5GPT-6 Astra1.05MOpenAIUnited StatesClosedSep 2026128,000
5Step 5 Preview1MStepFunChinaOpenSep 20261M
5MiMo-V2.6-Pro1MXiaomiChinaOpenSep 2026—
5Mistral Large 4 (preview)1MMistral AIFranceClosedOct 2026—
5Gemini 4 Argon1.05MGoogleUnited StatesClosedOct 202665,536
5Beam 1.01MReflection AIUnited StatesOpenOct 2026—

What's a context window? It's how much an AI model can take in at once: your question, any documents or code you paste in, the conversation so far and, for most models, the reply it writes. It's measured in tokens, the word fragments models read and write. By Anthropic's count, 1 million tokens is roughly 555,000 English words, so a 10-million-token window could in principle hold about ten times that. A bigger window doesn't mean the model makes good use of all of it. When Nvidia researchers tested long-context models with their RULER benchmark in 2024, only half of the models that claimed 32,000 tokens or more still performed well at 32,000. Very long prompts also cost more and take longer to answer.[10][13]

Basis and caveats: each lab's longest-context language model that the public can use today, either generally available or in public preview through the lab's own service, or downloadable as open weights. Models are ranked by the maximum context window stated in the lab's own documentation or model card, as of October 18, 2026. One row per lab, so a lab's other models at the same length aren't listed separately; where several tie, the row shows the lab's top model. Labs write “1 million” two ways, as 1,000,000 or as 1,048,576 (220), and OpenAI states 1,050,000. The difference is under 5%, so every model at about one million shares rank #5. The table shows each lab's own figure, “1M” or “1.05M”, and sorts by the exact number. Ties are listed in order of release. Most labs count the window as input and output together; OpenAI caps GPT-6 Astra's input at 922,000 tokens. These are the lengths the labs advertise, not measured ones: Llama 4 Scout was trained on 256,000-token sequences and MiniMax-Text-01 on 1-million-token ones before being stretched further, Qwen-Long reaches 10 million only through file uploads, and Pokee-Isaac 28B's results haven't been independently checked. Left out: models that were announced but never offered publicly (Magic's LTM-2-mini, described in 2024 with a 100-million-token window); models that have been withdrawn (Google's Gemini 1.5 Pro and xAI's Grok 4 Fast, both 2 million); extensions of another lab's model (such as Nvidia's 4-million-token UltraLong version of Meta's Llama 3.1 8B); services that route each request to other labs' models (Sakana AI's Fugu); and Meituan's LongCat-2.0, trained on 1-million-token data but published with a configuration that stops at 262,144 tokens. Mistral Large 4 counts as closed until Mistral publishes its weights, which it has promised for late October. Among labs not listed, LG's K-EXAONE 2.0 accepts 262,144 tokens. Limits change with nearly every release, so the order will too.[14][15][16]

The Top 10

Sixteen labs tie at #5 with about 1 million tokens. The six best known are written up below; the other ten are in the table.

#1

Qwen-Long (Alibaba) — 10 million tokens

Qwen-Long is the long-document model in Alibaba Cloud's Qwen lineup. Alibaba's Model Studio documentation lists a context window of 10,000,000 tokens, which Alibaba puts at about 15 million characters, or roughly 15,000 pages. There's a catch in how you reach it: a direct HTTP request can carry up to 1 million tokens, so anything longer has to be uploaded as files for the model to read. Replies are capped at 8,192 tokens, and Alibaba lists its prices and limits for the China (Beijing) region only. Alibaba now files it among its legacy models; its newer Qwen3.8 models stop at 1 million. Qwen-Long was already on sale in May 2024, when Alibaba cut its price by 97%, to 0.5 yuan per million input tokens.[1][2]

#1

Llama 4 Scout (Meta) — 10 million

Meta released Llama 4 Scout on April 5, 2025, with what it called an industry-leading 10-million-token context window, up from 128,000 in Llama 3. Scout is a mixture-of-experts model with about 109 billion parameters, 17 billion of them active for each token, and Meta says it fits on a single Nvidia H100 chip when compressed to 4-bit precision. It was pre-trained and post-trained on 256,000-token sequences, then stretched to 10 million in a later “mid-training” stage, using a design Meta calls iRoPE that drops position information from some layers so the model copes with longer inputs. Meta backed the figure with needle-in-a-haystack retrieval tests but published no detailed long-context scores. The weights are downloadable under the Llama 4 Community License; its bigger sibling, Llama 4 Maverick, stops at 1 million.[3]

#1

Pokee-Isaac 28B (Pokee AI) — 10 million

The newest model at 10 million tokens comes from Pokee AI, a start-up in Bellevue, Washington, founded by Bill Zhu, who previously led an applied reinforcement-learning group at Meta. Pokee-Isaac 28B appeared in early August 2026 as a preview served through Pokee's own API, at $0.15 per million input tokens and $1 per million output tokens, so filling the whole window once costs about $1.50. Replies are capped at 60,000 tokens. The weights aren't public. Pokee says the model uses a new, non-standard architecture, built partly on Alibaba's open Qwen3.6-27B, and can run on a single RTX 4090 graphics card. It reports a score of 93.3% on the RULER long-context test at 10 million tokens, but all its results are self-reported, and no independent lab had reproduced them as of mid-October 2026.[4][5]

#4

MiniMax-Text-01 (MiniMax) — 4 million

Shanghai-based MiniMax published MiniMax-Text-01 in January 2025 as an open-weight model with 456 billion parameters, 45.9 billion of them active for each token. Its model card says it was trained on contexts of up to 1 million tokens and can handle up to 4 million when generating. The trick is “lightning attention”, a form of linear attention whose cost grows roughly in step with the length of the text rather than with its square, mixed with some layers of standard attention. The 4-million figure applies when you run the downloadable weights yourself: MiniMax's own API no longer lists the model, and its current flagship, MiniMax-M3 (June 2026), offers 1 million tokens.[6]

#5

DeepSeek-V4-Pro (DeepSeek) — 1 million

DeepSeek's V4 preview, released on April 24, 2026, brought the Chinese lab's models to a 1-million-token context window. The larger of its two models, V4-Pro, has 1.6 trillion parameters, and its weights are free to download under the MIT license. DeepSeek's API documentation lists 1 million tokens of context and up to 384,000 tokens of output for both V4-Pro and the smaller V4.1-Flash, one of the most generous output limits in this table. DeepSeek updated V4-Pro in August 2026; that version (0813) is the one its API serves.[7]

#5

Grok 4.3 (xAI) — 1 million

xAI's Grok 4.3 reached the API on April 30, 2026, with a 1,000,000-token context window, the longest xAI currently lists, shared with its earlier Grok 4.20 models. Long prompts cost more: once a prompt passes 200,000 tokens, the whole request is billed at twice the normal rate. xAI used to go further. Grok 4 Fast, announced in September 2025, offered 2 million tokens, as did Grok 4.1 Fast that November, but neither appears in xAI's current model list, and its newest models, Grok 4.5 to Grok 4.7, are listed at 500,000.[8]

#5

Kimi K3 (Moonshot AI) — 1.05 million

Moonshot AI launched Kimi K3 in its apps and API on July 16, 2026, and published the weights at the end of the month. Its model card gives a context length of 1,048,576 tokens, and the model reads images and video as well as text. At 2.8 trillion parameters it is also the largest open-weight model RList ranks. For one web-browsing benchmark, Moonshot reports results both with “context compaction”, in which the model condenses its older history once it passes 300,000 tokens, and with the full 1-million-token window and no such help.[9]

#5

Claude Fable 5.1 (Anthropic) — 1 million

Anthropic released Claude Fable 5.1, its top model for demanding reasoning and long agentic tasks, on September 1, 2026. Like every model in Anthropic's current lineup (Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5), it has a 1-million-token context window and can write up to 128,000 tokens in a reply. Anthropic's documentation puts the window in plain terms: with the tokenizer introduced in Claude Opus 4.7, 1 million tokens is roughly 555,000 English words, or 2.5 million characters. A sister model with the same specifications, Claude Mythos 5.1, is reserved for organizations that pass Anthropic's verification programs.[10]

#5

GPT-6 Astra (OpenAI) — 1.05 million

OpenAI released GPT-6 Astra on September 3, 2026, first to enterprise security partners and then to ChatGPT subscribers and the API. Its model page lists a 1,050,000-token context window and up to 128,000 tokens of output, the same as the cheaper GPT-6.1 Sol and GPT-6 Luna. The window is shared between what goes in and what comes out, and OpenAI caps the input at 922,000 tokens to leave room for the reply. Prompts longer than 272,000 tokens are billed at twice the normal input rate. OpenAI first reached a million tokens with GPT-4.1 in April 2025, up from 128,000 in GPT-4o.[11]

#5

Gemini 4 Argon (Google) — 1.05 million

Google's newest frontier model was shown to a restricted group on September 30, 2026, and opened to developers through the Gemini API on October 13. Its documentation lists a context window of 1,048,576 tokens, shared by the input, the model's thinking and its output. Replies normally stop at 65,536 tokens, but a “long decode” mode lets Gemini 4 Argon write up to 1,048,576 tokens in a single response, the largest output allowance among these models. Google once went further: in June 2024 it opened a 2-million-token window on Gemini 1.5 Pro to all developers, but that model is no longer in Google's model list.[12]

In the spotlight Asia USA China South Korea France SpaceX Amazon Google
AI Language Models Context Windows Technology

Sources

  1. Alibaba Cloud Model Studio, Qwen-Long — help.aliyun.com; Model Studio text generation models (Qwen-Long listed as legacy, 10M; Qwen3.8 models 1M) — help.aliyun.com
  2. Yicai Global, "Alibaba Cloud Slashes Main AI Model Price by 97% to World's Lowest" (May 2024) — yicaiglobal.com
  3. Meta, "The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation" (April 5, 2025) — ai.meta.com; Llama 4 Scout model card — huggingface.co/meta-llama
  4. Pokee AI, Pokee-Isaac 28B model page — console.pokee.ai; OrcaRouter, "Pokee-Isaac 28B: Inside the 10M-Token Context Claim" — orcarouter.ai; MarkTechPost, "Pokee AI Releases Pokee-Isaac 28B" (Aug 8, 2026) — marktechpost.com
  5. GeekWire, "Point72 Ventures leads $12M seed round for a new AI agent startup founded by ex-Meta manager" — geekwire.com
  6. MiniMax, MiniMax-Text-01 model card — huggingface.co/MiniMaxAI; "MiniMax-01: Scaling Foundation Models with Lightning Attention" (arXiv 2501.08313) — arxiv.org; MiniMax API models — platform.minimax.io
  7. DeepSeek API docs, models and pricing — api-docs.deepseek.com; DeepSeek-V4-Pro-0813 model card — huggingface.co/deepseek-ai
  8. xAI, models and Grok 4.3 docs — docs.x.ai; xAI, "Grok 4 Fast" (Sept 19, 2025) — x.ai; OpenRouter, Grok 4.3 — openrouter.ai
  9. Moonshot AI, Kimi K3 model card — huggingface.co/moonshotai/Kimi-K3
  10. Anthropic, Claude Fable 5.1 — platform.claude.com; Anthropic, Models overview — platform.claude.com
  11. OpenAI, GPT-6 Astra model page — developers.openai.com; DataNorth, "OpenAI launches GPT-6 Astra" — datanorth.ai; OpenAI, "Introducing GPT-4.1 in the API" — openai.com
  12. Google, Gemini 4 Argon model page — ai.google.dev; Let's Data Science, "Google expands Gemini 4 Argon access across API and subscriber products" — letsdatascience.com; Google for Developers, "Gemini 1.5 Pro 2M context window, code execution capabilities, and Gemma 2 are available today" (June 27, 2024) — developers.googleblog.com
  13. Hsieh et al. (Nvidia), "RULER: What's the Real Context Size of Your Long-Context Language Models?" (arXiv 2404.06654) — arxiv.org
  14. Magic, "100M Token Context Windows" (LTM-2-mini, August 2024) — magic.dev; Nvidia, Llama-3.1-Nemotron-8B-UltraLong-4M-Instruct — huggingface.co/nvidia; eesel AI, "LongCat 2.0: inside Meituan's 1.6T open-weight model" — eesel.ai; DataNorth, "Sakana AI launches Fugu and Fugu Ultra" — datanorth.ai
  15. Other #5 rows, from each lab's documentation or model card: Amazon Nova 2 (docs.aws.amazon.com); Ant Group Ling-2.5-1T (huggingface.co/inclusionAI); Nvidia Nemotron 3 Ultra (huggingface.co/nvidia); Upstage Solar Open 2 (huggingface.co/upstage, koreatimes.co.kr); Z.ai GLM-5.3 (docs.z.ai, huggingface.co/zai-org); Tencent Hy4 preview (huggingface.co/tencent); StepFun Step 5 Preview (huggingface.co/stepfun-ai); Xiaomi MiMo-V2.6-Pro (huggingface.co/XiaomiMiMo); Mistral Large 4 (docs.mistral.ai, letsdatascience.com); Reflection AI Beam 1.0 (huggingface.co/reflection)
  16. BenchLM, "LLM Context Window Statistics" (cross-check of published limits, Oct 17, 2026) — benchlm.ai