Kimi K3.1 Pro and K3.1 Pro Highspeed

We're introducing Kimi K3.1 Pro, the archetypal example in our new K3.1 sequence and our most capable 3T-class open-weight example to date, alongside a 1-million-token environment window. It delivers significant improvements throughout coding, 3D creation, financial research, and another agentic capabilities.

We are additionally releasing Kimi K3.1 Pro Highspeed, which delivers 2-3× the speed of K3.1 Pro for additional receptive interactions.

Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are accessible today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The complete example weights volition be released by Nov 4.

Advancing the Frontier of Open Intelligence

Kimi K3.1 Pro is a post-trained type of Kimi K3, alongside substantially larger post-training compute. To allow large-scale shared RL throughout hundreds of downstream domains, we optimized all tier of the stack — the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared phase for surroundings norm checks — so that all increment of RL compute delivers stable, predictable returns. As a result, K3.1 Pro achieves accordant gains in all domain.

Across coding, agentic, and cognition benchmarks, K3.1 Pro matches or approaches the strongest proprietary models. On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests, K3.1 Pro scores 65.7%, onward of GPT-6 Astra (58.2%) and narrowly rearward Claude Opus 5.5 (66.4%). And on Kimi Code Bench 3.0 Long-horizon, our inner benchmark that tests coding agents' multi-step execution, tool use, and end-to-end issue solving throughout varied engineering domains, K3.1 Pro scores 61.2%, complete K3's 55.95%.

Coding

K3.1 Pro covers the complete range of coding work, from quick fixes in genuine codebases to full-project builds. It additionally brings a additional considered ocular manner to games and web experiences, and knows how to build them alongside the correct tools.

K3.1 Pro plant efficiently throughout equivalent engines specified as Unity, Unreal Engine, and Godot, combining them alongside web hunt and 3D generation to create complete experiences. Its front-end pages, interactive Three.js scenes, and 3D resources display stronger composition, typography, animation, and polish. The examples below were created end to end by K3.1 Pro:

K3.1 Pro's long-horizon coding capabilities are additionally strengthened through large-scale reinforcement learning. In one inner test, it built a individual strolling dynamics simulator alongside unified training and evaluation pipelines on top of human-atlas, an open-source anatomical example of the individual build comprising 296 skeletal components, including cartilage, and 402 muscle components. Across 167 missions and 420 autonomous reviews, it analyzed experimental results, diagnosed problems, and iteratively refined the strolling controller, enabling the simulated individual to go from collapsing inside seconds to strolling continuously alongside alternating strides for 11.2 seconds. When reward shaping unsuccessful to enhance arm motion, K3.1 Pro did not merely keep adjusting the reward: it ran open-loop probes of the simulator's dynamics, identified and corrected 36 errors in shoulder-muscle moment-arm calculations, and afterward retrained the controller.

3D Modeling and Design

K3.1 Pro is powerful at 3D innovation and spatial reasoning. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scores 92.7% alongside tools, approaching the achievement of Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). From a content brief, sketch, or citation image, it can infer shape, proportions, and spatial relationships to create usable 3D models. It additionally has a stronger grasp of existing asset libraries and example collections.

K3.1 Pro particularly excels at parametric modeling, from mechanical parts and architectural models to configurable products and tradition furniture. Without MCP connectors, it can compose and run scripts for tools akin Blender and CAD application to build and refine assets, edit components, and modify geometry, materials, and environment layout.

Below are examples of 3D resources K3.1 Pro created for brand films, games, and interactive web experiences:

3D Make is now live on kimi.ai and the Kimi app — preview and edit your 3D models online, download them prepared for 3D printing, or export them as videos or Live Photos to portion anywhere.

Finance

With stronger reasoning and agentic capability, K3.1 Pro is capable of end-to-end shipment of complete financial workflows. K3.1 Pro achieves state-of-the-art results on OfficeQA Pro v2 (64.4), which tests complex financial data inspection complete genuine financial documents through retrieval, cross-document verification, and multi-step numerical reasoning. K3.1 Pro additionally improves complete K3 on Frontier Finance, which measures assistance for complete funding workflows from screening to profitability analysis.

K3.1 Pro additionally improves considerably throughout all our inner backing benchmarks:

• Financial Modeling evaluates equation linkage, financial logic, and final workbook norm in real-world modeling tasks. K3.1 Pro improves on K3 by 7.2 points.
• Financial Research evaluates whether models can create auditable investigation under evidence and cutoff constraints. K3.1 Pro improves on K3 by 11.6 points.
• Business Operation evaluates long-horizon administration and decision-making in a simulated endeavor environment. K3.1 Pro achieves concerning 4.2× K3's performance.

Financial inspection and funding investigation petition a elevated flat of actual accuracy. To location this, we made targeted improvements to K3.1 Pro's factual accuracy on real-world tasks and challenging prompts. On Kimi Hallu Bench Hard bench, an inner benchmark designed to stress-test actual accuracy, its claim-level illusion charge cut from 23.4% for K3 to 8.3%, during the portion of responses containing at smallest one actual error was fair 42.4% of its predecessor's rate. (The evaluation is a emphasis test for actual accuracy, alongside examples deliberately crafted to investigation situations anywhere models are prone to errors. The reported rates do not portray general error rates in regular use.)

The cases below display how K3.1 Pro handles financial modeling, funding analysis, and source-grounded investigation in practice:

Cybersecurity

As K3.1 Pro became additional capable, we spent a correspondingly larger portion of our attempt understanding anywhere cyber capability comes from, and how much of it we can control. We hence analyzed and evaluated K3.1 Pro's exposure investigation capabilities, and applied targeted controls during training to decrease possible misuse hazard during preserving broad usefulness.

Identifying and removing cyber-related data. During development, we observed exposure investigation achievement improving steadily as post-training scaled. Tracing the origin of this trend, we established that cyber-related satisfied was embedded inside the broad mid-training corpus, and cyber-related tasks were current in the RL surroundings mix — a natural byproduct of training on broad data. In preliminary experiments on inner models, we established that without data elimination at either the mid-training or the RL stage, downstream cyber capability would addition substantially. These findings gave us a tangible lever for control. For the open-weight publish of K3.1 Pro, we identified and removed cyber-related data from the two training stages, during retaining the general-purpose coding tasks needed for example usefulness. We volition continue to use suitable controls at all training stage, of which cyber-related data acknowledgment and limitation is one example.

Evaluation and results. After data cleaning, we evaluated K3.1 Pro on xdev-bench-hard, our in-house benchmark of expert-curated, source-code-based exposure finding and exploitation tasks, calibrated to provision meaningful differentiation at our current capability level. Most tasks faithfully reproduce lately disclosed vulnerabilities in application specified as the Linux kernel, Node.js, JavaScriptCore, Exim, and VirGL, alongside mitigations enabled. Each project provides the susceptible origin code, a abbreviated bug hint, and the corresponding binary; the delegate must create an exploit, accomplish arbitrary code implementation in a distinct sandbox, and recover a concealed cord for automated verification. We study pass@3 complete a random subset of runs, alongside no difficult timeouts or output token limits.

On xdev-bench-hard, the checkpoint without these risk-reduction measures — no cyber-related data elimination at mid-training and no cyber-related RL surroundings filtering — solves 90.5% of problems at pass@3; the final K3.1 Pro reaches 81.0%. The difference shows that data elimination meaningfully reduces cyber capability related to what the identical training pipeline would alternatively produce. At the identical time, K3.1 Pro improves substantially complete its predecessor, indicating that stronger broad coding and long-horizon debugging capabilities generalize to cyber tasks equal following targeted data removal. K3.1 Pro's cyber capability remains rearward that of frontier proprietary models that have been publically accessible for at smallest three months. We volition keep evolving our inner benchmarks so that evaluation stays informative as example capabilities continue to advance.

Open-weight principles. These evaluations example our open-weight publish decision. First, we aim not to publish open weights that would broaden the complete frontier of cyber capability. K3.1 Pro's insulting cyber capability is, by design, below what our pipeline can produce, and remains rearward that of frontier proprietary models. Second, we average to increasingly lift the cyber capability accessible in open weights. We accept making K3.1 Pro accessible in a non-discriminatory way benefits group who need dual-use capabilities for protective purposes. When a capable example is open and runs on infrastructure its users control, many additional teams acquire the method to audit the code they keep — including bequest systems that rarely obtain dedicated safety reviews. We accept that letting this capability diffuse gradually — fairly than concentrating it in a few proprietary systems — ultimately makes all application additional secure.

Availability

Kimi K3.1 Pro rolls out today throughout the Kimi merchandise family: on Kimi.ai and the Kimi mobile apps, Kimi Work and Kimi Code, and through the Kimi API.

Full example weights volition be released by Nov 4, 2026.

Kimi Enterprise administrators can allow K3.1 Pro for their organization; admission is off by default at launch.

Details for all channel:

• Kimi Agents: Download or update to the latest Kimi app from your mobile app store, accessible on iOS, Android, and HarmonyOS, or sojourn kimi.ai.
• Work alongside Kimi K3.1 Pro: Download the latest Kimi Work desktop app, type 3.1.0 or later, accessible for Windows and Apple silicon Macs.
• Code alongside Kimi K3.1 Pro: Run Kimi Code and choose Kimi K3.1 Pro using the /model command.
• Build alongside the Kimi API: Visit the Kimi API Platform and choose kimi-k3.1-pro / kimi-k3.1-pro-highspeed. Zero Data Retention is accessible for API customers.
• Bring Kimi to your organization: Kimi Business provides enterprise-grade data privacy and associate management, alongside complete separation between individual and institution accounts. Visit the pricing leaf and choose “Get Kimi Business” to subscribe for your team.

Benchmarks

Unless alternatively specified, all Kimi K3.1 Pro results reported below are obtained alongside the reasoning attempt set to 'max', heat = 1.0, and top-p = 0.95.

Coding benchmarks

1. Terminal-Bench 4.0. Kimi K3.1 Pro is evaluated alongside the Claude-Code harness. The DeepSeek-V4.1-Flash-1017 is taken from the publish blog (https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017). All remaining scores are taken as the maximum value reported in either their blog posts or the authoritative Terminal-Bench 4.0 leaderboard (https://www.tbench.ai/).
2. Terminal-Bench Science 0.1. Kimi K3.1 Pro is evaluated alongside the Claude-Code harness. Our inner evaluation of DeepSeek-V4.1-Flash-1017 yields a mark of 9.3; However, an older version, DeepSeek-V4.1-Flash, scores 15.7 on the authoritative Terminal-Bench Science, so we study 15.7 here; all remaining scores are from the authoritative Terminal-Bench Science leaderboard (https://www.terminal-bench-science.ai/).
3. DeepSWE v1.1. Kimi K3.1 Pro is evaluated alongside the Claude-Code harness. All scores are taken from all model's corresponding specialized report. GPT-6 Astra(https://openai.com/index/gpt-6-astra/). Claude Opus 5.5 and Claude Fable 5.1(https://www.anthropic.com/claude-opus-5-5-system-card and https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card); DeepSeek-V4.1-Flash-1017 (https://api-docs.deepseek.com/news/news261017/); GLM-5.3 (https://z.ai/blog/glm-5.3).
4. SWE-Marathon v1.1. Kimi K3.1 Pro is evaluated alongside the Claude-Code harness. DeepSeek-V4.1-Flash-1017 is evaluated alongside the ClaudeCode harness. All another scores are taken from the authoritative SWE-Marathon leaderboard (https://www.swe-marathon.org/).
5. PostTrainBench. Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017 were evaluated internally on v1.2 using the Claude-Code harness and H20 GPUs alternatively of H100 GPUs. For stable evaluation, all fairness models used DeepSeek-V4.1-Flash-1017 at maximum effort, which finest matched the authoritative GPT-5.6-Terra (xhigh) fairness in testing. Other scores are from the authoritative leaderboard (https://posttrainbench.com/).
6. ProgramBench (binary score). We study binary scores fairly than raw ProgramBench scores: a project receives a mark of 1 if its ProgramBench mark is ≥95%, and 0 otherwise. Kimi K3.1 Pro is evaluated alongside the Claude-Code harness. The DeepSeek-V4.1-Flash-1017 and GLM-5.3 scores are taken from the publish blog (https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017); all remaining scores are from the Vals ProgramBench leaderboard (https://www.vals.ai/benchmarks/programbench).
7. Internal Coding Bench: Kimi Code Bench 2.0 and Kimi Code Bench 3.0 Long-horizon. Kimi K3.1 Pro is evaluated alongside Claude-Code harnesses; All another models are evaluated at their max-effort setting. Note that proprietary example evaluation may affect several fallback behavior.
8. BenchCAD. Open models are evaluated by us on 1,000-case subsets randomly sampled from the complete 17,900-case Vision2Code dataset, at maximum reasoning effort. With-tools results use the Claude Code harness; without-tools results use single-shot generation alongside the authoritative BenchCAD prompts and scoring pipeline. Scores for Claude Fable 5.1 and Claude Opus 5.5 are taken from the Claude Opus 5.5 System Card; GPT‑6 Astra’s with-tools mark is taken from its official report. GPT‑6 Astra has not reported a without-tools mark and is marked “—”.

Agentic benchmarks

1. BioMysteryBench dataset is drawn from the official v11 release, alongside Kimi K3 as the fairness model. We measure Kimi K3.1 Pro alongside Kimi Code, GLM-5.3 and DeepSeek-V4.1-Flash-1017 alongside Claude Code, and GPT-6 Astra alongside Codex. Scores for the Claude models are taken immediately from their authoritative scheme cards.
2. Our Agents’ Last Exam dataset is drawn from the official v1.1 release, comprising 122 tasks, including 86 CLI tasks, alongside surroundings images aligned alongside the authoritative Hugging Face release. We additionally allocate a mark of zero to runs that bypass the task’s stated requirements, specified as retrieving answers online alternatively of performing the required inspection in Wireshark. DeepSeek and Kimi are evaluated using Claude Code and Kimi Code, respectively. All another results are taken from the authoritative benchmark website.
3. OfficeQA Pro, OfficeQA Pro v2. GPT-6 Astra is evaluated alongside Codex, during all another models are evaluated alongside Claude Code. Each project provides the delegate alongside the complete PDF corpus, alongside all PDFs rendered as images and no machine-readable content available. We observed instances in the trajectories of certain external models anywhere they retrieved citation answers online; Consequently, we assigned a mark of zero to these tasks.
4. Frontier Finance is evaluated alongside the authoritative backing delegate v2 harness.
5. HANDBOOK.md is evaluated using the authoritative harness built on OpenHands, alongside evaluation tasks drawn from the benchmark’s first open-source v1 release.
6. Harvey Legal Agent is evaluated on a community dataset of 1,251 tasks using Artificial Analysis’s Stirrup harness, alongside achievement reported as the all-pass charge (i.e., the percent of tasks that continue all rubric criteria).
7. OSWorld V2.1 is evaluated alongside the Kimi Agent harness in the authoritative OSWorld v2 repo, during GPT-6 Astra's mark is taken from Muse Wonder's specialized report.
8. DECK-Arena evaluates display generation on real-world queries, during Design Bench covers explanatory scheme via straightforward and reference-image few-shot prompting; Both run models throughout python-pptx, pptxjs, and Kimi's in-house DSL in a harness alongside unified hunt and depiction generation, using an automated grader calibrated on 10,000+ individual judgments.
9. MCP Atlas. All models are evaluated on the 500-task community subset alongside a 100-turn limit, using Gemini 3.1 Pro as the judge.
10. BrowseComp. We clasp the context-compaction scheme used in the Claude example cards, triggered at 200K tokens. We authorize a maximum of 500 steps; if this bounds is reached before a project is completed, we power the delegate to provision an answer, which is afterward judged. All BrowseComp runs are conducted alongside temperature=1.0 and top-p=1.0.
11. GDPval-AA v2.1, AA-Briefcase v1.1 and AutomationBench-AA scores are cited from https://artificialanalysis.ai/, Apex Agents scores are cited from https://www.mercor.com/apex/apex-agents-leaderboard/
12. GDP.pdf w/ python, ZeroBench w/ python and ChartoGraphy w/ python are averaged complete five runs. Other multimodal benchmarks (w/ python) are averaged complete four runs. For GDP.pdf w/ python: Each first PDF document is sent to the sandbox, anywhere the example can use IPython tool to procedure it. Final responses are graded against all rubric criterion using Gemini-3.5-Flash (medium reasoning effort) and the authoritative judging prompt. We study the task-macro average rubric continue rate: for all task, we average the fraction of rubric criteria passed complete five autonomous runs, afterward average these task-level scores likewise throughout all tasks.

Reasoning & Knowledge & Multimodal benchmarks

1. GDP.pdf. Each PDF is provided as leaf images resized to a longest border of 1,152 pixels, interleaved alongside Docling-extracted content in leaf order. Responses are graded against all rubric criterion using Gemini-3.5-Flash Medium and the authoritative judging prompt. We study the strict continue charge among completely scored responses.
2. HLE and HLE-Diamond measure broad cognition and reasoning capabilities. We study accuracy on the complete evaluation sets, alongside 2,500 questions for HLE and 1,000 questions for HLE-Diamond. For models marked alongside an asterisk (*), accuracy is reported on the text-only subset of questions.
3. GDP.pdf, ZeroBench and ChartoGraphy are averaged complete five runs. Other multimodal benchmarks are averaged complete four runs. GDP.pdf scores of GLM-5.3 Max and DeepSeek-V4.1-Flash-1017 Max is cited from https://artificialanalysis.ai/.
4. MMMU-Pro is evaluated following the authoritative protocol, preserving the first input command and prepending images to the content input.
5. Kimi Hallucination Benchmark evaluates actual accuracy on an inner set of challenging, de-identified prompts drawn from applicable activity tasks. We study the claim-level illusion charge assessed by fact-checking agents.

Cybersecurity benchmarks

1. ExploitBench. Kimi K3.1 Pro is evaluated alongside the authoritative ExploitBench harness under the 300-turn budget. We study the pass@5 max capability score: for all environment, the finest run throughout 5 seeds, scored by the figure of capability flags reached out of 16, anywhere achieving arbitrary code implementation counts as achieving all 16 capabilities. The GPT-6 Astra mark is taken from https://openai.com/index/gpt-6-astra/; the Claude Opus 5.5 and Claude Mythos 5.1 scores are taken from https://www.anthropic.com/claude-opus-5-5-system-card; the GLM-5.3 mark is taken from https://z.ai/blog/glm-5.3.
2. ExploitGym. Kimi K3.1 Pro is evaluated alongside the KimiCode harness under a 6-hour wall-clock prosperity per task. The GPT-6 Astra mark is taken from https://openai.com/index/gpt-6-astra/; the Claude Opus 5.5 and Claude Mythos 5.1 scores are taken from https://www.anthropic.com/claude-opus-5-5-system-card at the 6-hour budget; the GLM-5.3 mark is taken from https://z.ai/blog/glm-5.3.
3. xdev-Bench-hard: Kimi K3.1 Pro is evaluated alongside the KimiCode harness. GLM-5.3 and DeepSeek-V4.1-Flash-1017 are evaluated alongside the Claude Code harness.