Glonce

Kimi K3.1 Pro open weights are out on Hugging Face, with a 1M-token context

modelsThe Glonce editorsHacker News🔥85Glonce rating85Impact80Novelty65Relevance100Surprise55Players60Credibility85
Kimi K3.1 Pro open weights are out on Hugging Face, with a 1M-token context

The moonshotai organization on Hugging Face has published open weights and the model card for Kimi K3.1 Pro. The card calls it an open-weight, natively multimodal agentic model and the developers' most capable "3T-class" model to date. No exact parameter count is given beyond that label.

Kimi K3.1 Pro is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and it is the first model in the K3.1 series. It has a 1-million-token context window. The developers say it delivers substantial improvements in coding, 3D creation, financial research, factual accuracy and other agentic capabilities. The visible text gives no figures behind that claim.

The one architecture change the card lists against Kimi K3: the full-attention Gated MLA layers are replaced with Gated DeepSeek Sparse Attention (Gated DSA) layers, for more efficient long-context modeling.

The card has a long evaluation section. Unless stated otherwise, all Kimi K3.1 Pro results use reasoning effort max, temperature 1.0 and top-p 0.95. The footnotes show how wide the test suite is. Coding: Terminal-Bench 4.0, Terminal-Bench-Science 0.1, SWE-Marathon v1.1, ProgramBench, DeepSWE v1.1, PostTrainBench, two internal Kimi Code Bench sets and BenchCAD. Agentic work: BioMysteryBench, Agents' Last Exam, OfficeQA Pro, Frontier Finance, HANDBOOK.md, Harvey Legal Agent, OSWorld V2.1, DECK-Arena and Design Bench, BrowseComp, GDPval-AA v2.1 and APEX-Agents. Reasoning, knowledge and multimodal: GDP.pdf, HLE-Full and HLE-Diamond, MMMU-Pro, PerceptionBench and an internal Kimi Hallucination Benchmark. Cybersecurity: ExploitBench, ExploitGym and an internal xdev-bench-hard. Most coding runs for K3.1 Pro use the Claude Code harness; some agentic and cybersecurity runs use Kimi Code. The benchmark score tables themselves are not in the text available here, so no claim of beating or trailing any named model can be made.

Several footnotes explain how comparisons were built. On Terminal-Bench-Science 0.1, the developers' internal evaluation of DeepSeek-V4.1-Flash-1017 gave 9.3, but the table instead reports 15.7 for the older DeepSeek-V4.1-Flash from the official leaderboard. ProgramBench is reported as binary: a task scores 1 if its ProgramBench score is at least 95%. BrowseComp uses context compaction at 200K tokens and a 500-step maximum. ExploitBench uses a 300-turn budget and reports pass@5 maximum capability; ExploitGym gives each task a six-hour wall-clock budget, and Claude comparison results use the same budget. Harvey Legal Agent was run on 1,251 public tasks and reports an all-pass rate. Agents' Last Exam v1.1 has 122 tasks, 86 of them CLI tasks. HLE-Full has 2,500 questions and HLE-Diamond 1,000. Kimi Code Bench 2.0 and 3.0 Long-horizon are internal benchmarks, and the footnote warns that proprietary-model testing may involve fallback behavior.

On access: the API is on https://platform.kimi.ai under the model name kimi-k3.1-pro, with OpenAI- and Anthropic-compatible endpoints. The thinking effort is set with the top-level reasoning_effort request field, which accepts "low", "high" and "max", with "max" as the default. For multi-turn conversations and tool calls, the model requires the complete assistant message returned by the API to be passed back as-is, including reasoning_content and tool_calls, not just content. The Hugging Face page also carries ready snippets for Transformers (with trust_remote_code), vLLM, SGLang and Docker Model Runner.

Key facts

  • Kimi K3.1 Pro is published as open weights under the moonshotai namespace on Hugging Face; the card calls it a natively multimodal agentic model in the "3T-class", with no exact parameter count.
  • It is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and has a 1-million-token context window.
  • Versus Kimi K3, full-attention Gated MLA layers are replaced with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling.
  • The API is on platform.kimi.ai as kimi-k3.1-pro, with OpenAI/Anthropic-compatible endpoints; reasoning_effort accepts low, high or max (default max), and the full assistant message must be passed back as-is.
  • The evaluation covers coding, agentic, multimodal and cybersecurity benchmarks, but the score tables are not in the text available here, only the footnotes.

Why it matters

This is a large open-weight release aimed at agentic work: coding, financial research, 3D creation and computer use, with a 1-million-token context window. The model is described as the first of a K3.1 series, post-trained from Kimi K3. The one architectural change listed is the swap of full-attention Gated MLA layers for Gated DeepSeek Sparse Attention layers, aimed at cheaper long-context handling. The claimed gains in coding, 3D creation, financial research and factual accuracy come from the developers' own card.

Who it affects

Developers building agents and coding tools who want an open-weight model or an API that mimics the OpenAI and Anthropic formats. Teams doing long-document work, such as finance, legal or office-document tasks, will notice that the benchmark list includes Harvey Legal Agent, Frontier Finance and OfficeQA Pro. Researchers who track cybersecurity capability evaluations will see ExploitBench, ExploitGym and an internal xdev-bench-hard in the list.

How to use it

Through the API: use https://platform.kimi.ai and select the model name kimi-k3.1-pro. Set effort with the top-level reasoning_effort field (low, high or max; the default is max). In multi-turn chats and tool calls, send back the full assistant message exactly as the API returned it, including reasoning_content and tool_calls. With the weights: the Hugging Face page shows snippets for Transformers (trust_remote_code=True), vLLM, SGLang and Docker Model Runner, using an OpenAI-compatible server and image-plus-text messages. The list of recommended inference engines is empty in the text available here. No API pricing, license or hardware requirements appear in it either.

How solid is it

The material is the developers' own model card, so every capability claim is theirs and none is independently checked. The text available here is truncated and the benchmark result tables are absent, so there are no scores to weigh. The footnotes do show the setup: most K3.1 Pro coding runs use the Claude Code harness, with maximum reasoning effort and temperature 1.0. Some benchmarks are internal (Kimi Code Bench 2.0 and 3.0 Long-horizon, the Kimi Hallucination Benchmark, xdev-bench-hard), and some comparison scores come from other vendors' reports or leaderboards rather than from the developers' own runs.

Risks and caveats

Comparisons mix harnesses and sources: some rivals' scores are taken from release blogs, system cards or leaderboards, and a footnote shows the table using a leaderboard figure (15.7) for an older DeepSeek-V4.1-Flash instead of the developers' internal 9.3 for DeepSeek-V4.1-Flash-1017. Internal benchmarks cannot be reproduced from the card alone. Hardware needs for running the weights locally are not given, and neither are the license or API pricing, so check the full repository before planning a deployment. Agents that drop reasoning_content or tool_calls when replaying history would break the stated requirement for multi-turn use.

“Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date.”

— Kimi K3.1 Pro model card, Hugging Face