Kimi K3.1 Pro: Open-Weight Model With 1M Context

Kimi K3.1 Pro open-weight multimodal agentic model from Moonshot AI

Moonshot AI has released Kimi K3.1 Pro, an open-weight, natively multimodal agentic model with 2.8 trillion parameters. The company calls it its most capable model in the 3T class so far. It has a 1-million-token context window and is the first model in the K3.1 series.

The weights are on Hugging Face, and the model is also available through Moonshot’s API.

What changed from Kimi K3

Kimi K3.1 Pro is post-trained from Kimi K3. Moonshot used large-scale joint reinforcement learning across hundreds of downstream domains. The company says that brought substantial improvements in coding, 3D creation, financial research and factual accuracy, along with other agentic abilities.

One architectural change stands out. Moonshot replaced the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention layers, known as Gated DSA, to handle long contexts more efficiently. That matters for a model with a million-token window, because full attention gets expensive as inputs grow.

Benchmarks and how they were run

Moonshot reports results on coding, agentic, reasoning, multimodal and cybersecurity benchmarks. Unless noted, all results use the maximum reasoning effort, a temperature of 1.0 and a top-p of 0.95. Many coding evaluations ran inside Claude Code. Others used Kimi Code, Moonshot’s own tool.

The comparison set is large and recent. It includes GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1, DeepSeek-V4.1-Flash-1017 and GLM-5.3. Some competitor scores come from those labs’ reports or official leaderboards, which means test conditions aren’t always identical.

  • Coding: Terminal-Bench 4.0, SWE-Marathon v1.1, ProgramBench, DeepSWE v1.1 and BenchCAD, plus Moonshot’s internal Kimi Code Bench 2.0 and 3.0.
  • Agentic: BioMysteryBench, Agents’ Last Exam and OfficeQA Pro.
  • Reasoning and multimodal: HLE, MMMU-Pro, GDP.pdf and a Kimi Hallucination Benchmark that measures claim-level hallucination rates on prompts drawn from practical work tasks.
  • Cybersecurity: ExploitBench, ExploitGym and xdev-bench-hard.

Moonshot’s notes flag where scores are its own, so readers can see which numbers it produced itself. PostTrainBench was run internally on H20 GPUs instead of H100s, and the Kimi benchmarks are internal. Treat the numbers as the vendor’s claims until outside groups repeat them.

Test details worth knowing

Some setup choices shape how to read the scores. Agents’ Last Exam uses 122 tasks, 86 of them command-line tasks, and runs that bypass the stated requirements get zero. The full Humanity’s Last Exam set has 2,500 questions and the Diamond subset 1,000. GDP.pdf, ZeroBench and ChartoGraphy results are averaged over five runs, and other multimodal tests over four.

On cybersecurity, ExploitBench gave the model a 300-turn budget, while ExploitGym allowed six hours of wall-clock time per task. In ExploitBench, arbitrary code execution counts as reaching all 16 capability flags.

Using the model

The API is on platform.kimi.ai under the name kimi-k3.1-pro, and it works with OpenAI-style and Anthropic-style requests. A top-level reasoning_effort field sets how hard the model thinks. It accepts low, high and max, and max is the default.

Multi-turn chats and tool calls have a requirement developers should know about. The full assistant message returned by the API, including reasoning_content and tool_calls, must be passed back unchanged, with nothing dropped. The model relies on its earlier reasoning when it continues a conversation.

For running it yourself, the model page gives setup steps for Transformers, vLLM, SGLang and Docker Model Runner. Its documentation also covers vision input, structured output, tool choice, dynamic tool loading and context caching.

License and adoption

Kimi K3.1 Pro ships under the Kimi K3.1 Pro License, so anyone planning commercial use should read its terms first. The Hugging Face page shows 64,576 downloads in the last month.

What to watch next

A 2.8-trillion-parameter model needs serious hardware, so the first practical question is who can host it and at what price. Hosted access through the API gets around that for most teams.

The next checks are independent results. If outside leaderboards confirm Moonshot’s coding and agentic numbers against the other frontier systems on the list, K3.1 Pro would stand as one of the strongest open-weight options available.

Source: huggingface.co

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.