Kimi K3.1 Pro API

Kimi K3.1 Pro is an open-weight, natively multimodal agentic model developed by Moonshot AI, which the company describes as its most capable 3T-class model to date. It 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. With a 1-million-token context window, Moonshot AI reports substantial improvements in coding, 3D creation, financial research, factual accuracy and other agentic capabilities.

Moonshot AI reports that across coding, agentic and knowledge benchmarks K3.1 Pro matches or approaches the strongest proprietary models. On Terminal-Bench 4.0 it scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%). On Kimi Code Bench 3.0 Long-horizon, Moonshot AI's internal benchmark for multi-step execution, tool use and end-to-end problem solving by coding agents, it scores 61.2% where Kimi K3 scored 55.95%.

Key properties reported by Moonshot AI:

• Post-trained from Kimi K3 at a much larger scale. K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. To make joint reinforcement learning across hundreds of downstream domains work, Moonshot AI optimized every layer of the stack: the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared platform for environment quality checks. Each increment of RL compute delivers stable, predictable returns. Moonshot AI reports consistent gains in every domain.
• Coding, from quick fixes to full projects. K3.1 Pro covers the full range of coding work, from fixes in real codebases to full-project builds. It works across game engines such as Unity, Unreal Engine and Godot, combining them with web search and 3D generation to create complete experiences. Moonshot AI reports stronger composition, typography, animation and polish in its front-end pages, interactive Three.js scenes and 3D assets.
• Long-horizon engineering. In one internal test, K3.1 Pro built a human walking dynamics simulator with integrated training and evaluation pipelines on top of an open-source anatomical model with 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews it analyzed results, diagnosed problems and refined the walking controller until the simulated human walked continuously with alternating strides for 11.2 seconds. When reward shaping stopped improving arm motion, it probed the simulator's dynamics directly and corrected 36 errors in shoulder-muscle moment-arm calculations before retraining.
• 3D modeling and design. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). It is particularly strong at parametric modeling, from mechanical parts and architectural models to configurable products and custom furniture. It can write and run scripts for tools like Blender and CAD software to build assets, edit components and adjust geometry, materials and scene layout.
• Finance. K3.1 Pro reaches a state-of-the-art 64.4 on OfficeQA Pro v2, which tests fiscal data analysis over real financial documents through retrieval, cross-document checking and multi-step numerical reasoning. On Moonshot AI's internal finance benchmarks it improves on Kimi K3 by 7.2 points in Financial Modeling (57.6 vs 50.4) and by 11.6 points in Financial Research (70.4 vs 58.8). In Business Operation, a long-horizon management simulation, it reaches about 4.2 times the score of K3 (37 vs 8.9).
• Factual accuracy. On Kimi Hallu Bench Hard, an internal stress test for factual accuracy, the claim-level hallucination rate fell from 23.4% for Kimi K3 to 8.3%. The share of responses with at least one factual error was 42.4% of its predecessor's rate. Moonshot AI notes that the benchmark is built from prompts that deliberately probe error-prone situations, so these rates do not represent overall error rates in daily use.

Architecture

All architecture figures are from Moonshot AI's official model card.

Property	Value
Architecture	Mixture-of-Experts (MoE)
Total Parameters	2.8T
Activated Parameters	104B
Number of Layers	93
Number of Dense Layers	1
Attention-Layer Composition	69 KDA + 24 Gated DSA
Attention Hidden Dimension	7168
Number of Attention Heads	96
DSA Index Head Dimension	128
DSA Index Number of Heads	32
DSA Index Top-K	2048
Latent MoE Dimension	3584
MoE Hidden Dimension (per Expert)	3072
Number of Experts	896
Selected Experts per Token	16
Number of Shared Experts	2
Vocabulary Size	160K
Context Length	1048576
Attention Mechanism	KDA & Gated DSA
Activation Function	SiTU-GLU
Vision Encoder	MoonViT-V2
Parameters of Vision Encoder	401M
Quantization	MXFP4 weights / MXFP8 activations (quantization-aware training)
Modality	Text, Image

Compared with Kimi K3, Moonshot AI replaced the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling. The rest of the published architecture matches Kimi K3.

Benchmarks

All results below are from Moonshot AI's official model card. Unless otherwise specified, all Kimi K3.1 Pro results use reasoning effort max, temperature = 1.0 and top-p = 0.95. "n/a" means the source table does not report a score. Higher is better except where a benchmark name states lower is better. Elo scores and other benchmark-specific metrics are not directly comparable with accuracy or pass-rate scores.

Coding

Benchmark	Kimi K3.1 Pro (max)	Claude Fable 5.1 (max, w/ fallback)	GPT-6 Astra (max, w/ fallback)	Claude Opus 5.5 (max, w/ fallback)	DeepSeek-V4.1-Flash-1017 (max)	GLM-5.3 (max)
Terminal-Bench 4.0	65.7	57.9	58.2	66.4	42.3	41.8
Terminal-Bench-Science 0.1	38.1	40.0	68.1	63.3	15.7	8.1
SWE Marathon v1.1	51.7	45.6	42.5	46.3	40.0	42.5
BenchCAD w/o tools	72.8	60.6	n/a	73.0	41.2	n/a
BenchCAD w/ tools	92.7	92.6	95.9	96.2	81.6	n/a
Program Bench	31.5	53.5	50.0	65.0	29.1	19.0
DeepSWE v1.1	75.5	67.4	74.1	74.2	77.0	66.9
PostTrain Bench v1.2	36.5	44.6	41.9	43.8	30.5	34.5
Kimi Code Bench 2.0 (Internal)	77.1	n/a	n/a	n/a	76.1	61.7
Kimi Code Bench 3.0 Long-horizon (Internal)	61.2	58.8	66.5	58.5	51.5	45.5

Agentic

Benchmark	Kimi K3.1 Pro (max)	Claude Fable 5.1 (max, w/ fallback)	GPT-6 Astra (max, w/ fallback)	Claude Opus 5.5 (max, w/ fallback)	DeepSeek-V4.1-Flash-1017 (max)	GLM-5.3 (max)
HLE-Full w/ tools	70.0	65.6	57.2	67.7	58.8	62.5
BrowseComp	92.1	n/a	91.5	90.8	n/a	n/a
BioMysteryBench	80.6	81.6	78.9	81.9	80.0	79.6
CharXiv (RQ) w/ python	93.4	92.7	92.8	93.8	85.6	n/a
MMMU-Pro w/ python	90.0	88.4	87.8	89.8	82.5	n/a
GDP.pdf w/ python	85.7	83.8	85.5	83.9	73.0	n/a
ChartoGraphy w/ python	90.2	88.4	91.0	89.0	83.0	n/a
BabyVision w/ python	95.4	95.1	96.6	97.7	92.5	n/a
ZeroBench_main w/ python (pass@5)	63.0	59.0	58.0	61.0	51.0	n/a
Agents' Last Exam (ALL)	43.2	n/a	41.8	42.6	37.7	26.2
Agents' Last Exam (CLI)	44.2	n/a	44.2	40.7	36.0	29.4
Automation Bench-AA	73.3	59.4	68.5	69.5	66.6	62.2
GDPval-AA v2.1 (Elo-score)	1846	1754	1573	1866	1746	1650
AA-Briefcase v1.1 (Elo-score)	1678	1672	1569	1803	1570	1507
APEX-Agents	69.4	68.6	64.7	73.5	69.8	56.6
OfficeQA Pro v2	64.4	54.5	54.4	62.2	40.4	31.1
OfficeQA Pro	67.4	68.6	66.2	n/a	62.4	58.6
Frontier Finance	58.6	60.8	45.0	56.7	54.5	57.5
Harvey Legal Agent	23.7	19.1	n/a	n/a	21.0	n/a
HANDBOOK.md	44.6	38.8	n/a	41.5	28.9	20.8
OSWorld V2.1	79.6	80.7	75.8	81.8	69.3	n/a
Business Operation (Internal)	37.8	30.7	32.3	n/a	8.9	18.6
DECK-Arena (Internal, Elo-score)	1456	n/a	1423	1253	1231	1119
Design Bench (Internal, Elo-score)	957	n/a	1076	927	863	718
Kimi Hallu Bench Hard w/ tools (Internal, lower is better)	8.3	23.3	5.5	15.0	23.3	22.6

Reasoning, Knowledge and Multimodal

Benchmark	Kimi K3.1 Pro (max)	Claude Fable 5.1 (max, w/ fallback)	GPT-6 Astra (max, w/ fallback)	Claude Opus 5.5 (max, w/ fallback)	DeepSeek-V4.1-Flash-1017 (max)	GLM-5.3 (max)
GDP.pdf	34.6	27.6	34.2	30.6	17.2	11.2
HLE-Full	51.1	57.6	53.6	59.8	42.6	43.0*
HLE-Diamond	39.4	55.1	65.8	60.9	21.6	25.1*
GPQA-Diamond	94.2	93.7	96.1	93.9	92.2	91.7
CharXiv (RQ)	89.0	88.8	90.0	92.1	77.0	n/a
MMMU-Pro	84.1	86.3	86.9	87.7	80.6	n/a
ZeroBench_main (avg@5 / pass@5)	29.8 / 39.0	23.0 / 35.0	43.2 / 52.0	35.0 / 43.0	11.8 / 23.0	n/a
PerceptionBench	68.3	63.0	77.6	72.1	47.7	n/a
Kimi Hallu Bench w/o tools (Internal, lower is better)	5.4	11.4	4.2	6.8	13.1	14.8

Cyber Security

Benchmark	Kimi K3.1 Pro (max)	Claude Fable 5.1 (max, w/ fallback)	GPT-6 Astra (max, w/ fallback)	Claude Opus 5.5 (max, w/ fallback)	DeepSeek-V4.1-Flash-1017 (max)	GLM-5.3 (max)
ExploitBench	74.7	83.0	100	91.0	n/a	54.4
ExploitGym	20.7	30.4	42.4	34.5	n/a	15.0
xdev-bench-hard (Internal)	81.0	n/a	n/a	n/a	85.7	57.1

* Accuracy on the text-only subset of HLE.

Benchmark methodology notes (as disclosed by Moonshot AI)

• Terminal-Bench 4.0. Kimi K3.1 Pro is evaluated with the Claude Code harness. The DeepSeek-V4.1-Flash-1017 score is taken from its release blog. All remaining scores are the maximum values reported in model blog posts or on the official Terminal-Bench leaderboard.
• Terminal-Bench-Science 0.1. Kimi K3.1 Pro is evaluated with Claude Code. Moonshot AI's internal evaluation of DeepSeek-V4.1-Flash-1017 yields 9.3. The table instead reports 15.7 for the older DeepSeek-V4.1-Flash from the official leaderboard. All remaining scores are from the Terminal-Bench-Science leaderboard.
• SWE-Marathon v1.1. Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017 are evaluated with Claude Code. Other scores are from the official leaderboard.
• ProgramBench. The table reports binary scores: a task receives 1 if its ProgramBench score is at least 95% and 0 otherwise. K3.1 Pro is evaluated with Claude Code. DeepSeek and GLM scores are taken from the DeepSeek release blog. Remaining scores are from Vals AI.
• DeepSWE v1.1. K3.1 Pro is evaluated with Claude Code. Other scores come from each model's technical report.
• PostTrainBench. K3.1 Pro and DeepSeek-V4.1-Flash-1017 are evaluated internally on v1.2 with Claude Code and H20 GPUs rather than H100 GPUs. For stable evaluation all judge models use DeepSeek-V4.1-Flash-1017 at maximum effort, which best matched the official GPT-5.6-Terra (xhigh) judge in testing. Other scores are from the official leaderboard.
• Kimi Code Bench 2.0 and Kimi Code Bench 3.0 Long-horizon. These are internal benchmarks. K3.1 Pro is evaluated with Claude Code. All other models use their maximum-effort setting. Proprietary-model testing may involve fallback behavior.
• BenchCAD. Open models are evaluated on 1,000-case subsets randomly sampled from the 17,900-case Vision2Code dataset at maximum reasoning effort. With-tools results use Claude Code. Without-tools results use single-shot generation with the official prompts and scoring. Claude scores are from the Claude Opus 5.5 System Card and GPT-6 Astra's with-tools score is from its official report. GPT-6 Astra has no reported without-tools score.
• BioMysteryBench. Uses the official v11 release with Kimi K3 as judge. K3.1 Pro uses Kimi Code. GLM-5.3 and DeepSeek-V4.1-Flash-1017 use Claude Code. GPT-6 Astra uses Codex. Claude scores are from official system cards.
• Agents' Last Exam. Uses the official v1.1 release of 122 tasks, including 86 CLI tasks. Runs that bypass stated requirements receive zero. DeepSeek is evaluated with Claude Code and Kimi with Kimi Code. Other results are from the official benchmark website.
• OfficeQA Pro and OfficeQA Pro v2. GPT-6 Astra uses Codex. Other models use Claude Code. Each task provides the complete PDF corpus as page images without machine-readable text. External-model trajectories that retrieved reference answers online receive zero for those tasks.
• Frontier Finance. Evaluated with the official Finance Agent v2 harness.
• HANDBOOK.md. Evaluated with the official OpenHands-based harness on tasks from the benchmark's early open-source v1 release.
• Harvey Legal Agent. Evaluated on 1,251 public tasks with Artificial Analysis's Stirrup harness. The score is the all-pass rate: the percentage of tasks passing every rubric criterion.
• OSWorld V2.1. Kimi models use the Kimi Agent harness in the official OSWorld v2 repository. GPT-6 Astra's score is taken from Muse Wonder's technical report.
• DECK-Arena and Design Bench. DECK-Arena evaluates presentation generation on real-world queries. Design Bench covers graphic design through direct and reference-image few-shot prompting. Both evaluate python-pptx, pptxjs and Kimi's in-house DSL in a harness with integrated search and image generation. The automated grader is calibrated on more than 10,000 human judgments.
• BrowseComp. Uses context compaction triggered at 200K tokens with a 500-step maximum. If the limit is reached, the agent is forced to answer before judging. All runs use temperature = 1.0 and top-p = 1.0.
• GDPval-AA v2.1, AA-Briefcase v1.1 and AutomationBench-AA. Scores are cited from Artificial Analysis. APEX-Agents scores are cited from the official leaderboard.
• Tool-augmented multimodal evaluations. GDP.pdf w/ python, ZeroBench w/ python and ChartoGraphy w/ python are averaged over five runs. Other multimodal benchmarks with Python are averaged over four. For GDP.pdf w/ python each original PDF is sent to a sandbox for IPython processing and responses are graded against each rubric criterion by Gemini-3.5-Flash at medium reasoning effort with the official judging prompt. The score is the task-macro mean rubric pass rate.
• GDP.pdf. Each PDF is provided as page images resized to a longest edge of 1,152 pixels and interleaved with Docling-extracted text in page order. Responses are graded against each rubric criterion with the official judging prompt. The table reports the strict pass rate among fully scored responses. GLM-5.3 and DeepSeek-V4.1-Flash-1017 scores are cited from Artificial Analysis.
• HLE-Full and HLE-Diamond. The full evaluation sets contain 2,500 and 1,000 questions respectively.
• MMMU-Pro. Evaluated with the official protocol, preserving the original input order and prepending images to the text input.
• PerceptionBench. An in-house benchmark focused on atomic visual perception.
• Kimi Hallucination Benchmark. Evaluates factual accuracy on an internal set of challenging, de-identified prompts drawn from practical work tasks. The table reports claim-level hallucination rates assessed by fact-checking agents.
• ExploitBench. K3.1 Pro is evaluated with the official harness under a 300-turn budget. The table reports the pass@5 maximum-capability score: for each environment the best of five seeds is scored by the capability flags reached out of 16. Arbitrary code execution counts as all 16 capabilities.
• ExploitGym. K3.1 Pro is evaluated with Kimi Code under a six-hour wall-clock budget per task. Comparison scores are taken from official reports, with Claude results using the same six-hour budget.
• xdev-bench-hard. An internal benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks. K3.1 Pro is evaluated with Kimi Code. GLM-5.3 and DeepSeek-V4.1-Flash-1017 use Claude Code. The model is evaluated at pass@3 over a random subset of runs, without hard timeouts or output-token limits.

Cybersecurity and Responsible Release

Disclosed by Moonshot AI in its release announcement:

• Removing cyber-related training data. During development Moonshot AI observed vulnerability-research performance rising steadily as post-training scaled. It traced the trend to cyber-related content inside the general mid-training corpus and to cyber-related tasks in the RL environment mix. For the open-weight release of K3.1 Pro it identified and removed cyber-related data from both stages while keeping the general-purpose coding tasks the model needs to be useful.
• Measured effect. On xdev-bench-hard, a checkpoint trained without these risk-reduction measures solves 90.5% of problems at pass@3. The released K3.1 Pro reaches 81.0%. Moonshot AI reads this as data removal meaningfully reducing cyber capability, while stronger general coding and long-horizon debugging still carry over to cyber tasks. It states that K3.1 Pro's cyber capability remains behind frontier proprietary models that have been publicly available for at least three months.
• Open-weight principles. Moonshot AI aims not to release open weights that would extend the absolute frontier of cyber capability, and it intends to raise the cyber capability available in open weights progressively. It argues that broad access to capable open models helps defenders audit the code they maintain, including legacy systems that rarely receive dedicated security reviews.

Thinking and Tool Use on Run BiOS

• Reasoning is always on. Kimi K3.1 Pro always thinks before answering and returns its reasoning content alongside the response. Thinking effort is set with reasoning_effort at low, high or max. Moonshot AI's own default is max, the setting used for every benchmark above. On Run BiOS a request that does not set an effort runs at high.
• Preserved thinking. For multi-turn conversations and tool calls, Moonshot AI requires the complete assistant message to be passed back as returned, including its reasoning content and tool calls, not only the final text.
• Function calling. Structured tool use with multi-step tool calls for agentic workflows.
• Vision input. The model accepts image input natively through its MoonViT-V2 vision encoder.
• Structured outputs. JSON mode and JSON schema responses.
• Prompt caching. Repeated prefixes are served from cache at a reduced rate.

License

Kimi K3.1 Pro is released by Moonshot AI under the Kimi K3.1 Pro License.

On Run BiOS

Kimi K3.1 Pro is available on Run BiOS as the serverless model moonshotai/kimi-k3.1-pro, with a 1,048,576-token context window, up to 65,536 output tokens, always-on reasoning at effort low, high or max, image input, function calling, structured outputs and automatic prompt caching. Kimi K3 remains available as moonshotai/kimi-k3.