Kimi K3.1 Pro

Kimi K3.1 Pro is a multimodal large language model from the Chinese company Moonshot AI, announced on October 23, 2026. Moonshot describes it as the first model in a new K3.1 series and "our most capable 3T-class open-weight model to date."^[1][2] According to Moonshot, K3.1 Pro is a post-trained version of Kimi K3, the mixture-of-experts model it released in July 2026, trained with "substantially larger post-training compute."^[1][10] The company launched it alongside a faster variant, Kimi K3.1 Pro Highspeed, and made both available the same day in the Kimi apps, Kimi Work, Kimi Code, and the Kimi API.^[1][2]

Moonshot's launch post on X gives the model 2.8 trillion total parameters, 104 billion active parameters, and a 1-million-token context window.^[2] The company said the full model weights would be released "by Nov 4, 2026."^[1][2] As of October 25, 2026, the weights had not been published, and no license for them had been announced.^[1][12] Moonshot's benchmark results for the model are self-reported. They place it ahead of or close to Claude Opus 5.5 and GPT-6 Astra on several coding and agent tests and behind them on others, including most knowledge, perception and cybersecurity evaluations.^[1]

Release and availability

Moonshot announced the model on its X accounts on October 23, 2026 and published a technical blog post on kimi.ai the same day.^[1][2][3] The blog lists these launch channels:^[1]

The API is OpenAI-compatible. The API platform guide states that K3.1 Pro "is a flagship model": it is unlocked after an account top-up of at least US$1, and the cumulative top-up amount sets the account's rate-limit tier.^[4] The platform's model list describes kimi-k3.1-pro as "the first model of the new K3.1 series and Kimi's most capable flagship to date" and kimi-k3.1-pro-highspeed as having the "same intelligence, about 2-3× the speed."^[6] On the same page Moonshot now points users of retired models to K3.1 Pro. It also says that from October 31, 2026, K2 series models will no longer be offered to newly registered users, and that it plans to retire all K2 series models by the end of December 2026.^[6]

OpenRouter lists the model as moonshotai/kimi-k3.1-pro, with a release date of October 23, 2026, a 1,048,576-token context window, and text, image and video input.^[9]

API pricing

Moonshot's API pricing page listed the following rates per million tokens, excluding taxes, as of October 25, 2026:^[5]

The standard K3.1 Pro rates are therefore lower than K3's, while the Highspeed variant costs twice the standard K3.1 Pro rate for input, output and cache writes.^[5] Moonshot states that all K3 series models support a 1,048,576-token context window and that K3.1 Pro uses flat pricing with no tiering by context length.^[4][5] Context caching is automatic. Cache writes have a default time-to-live of five minutes, and a one-hour tier can be selected with the prompt_cache_options.ttl request field.^[4]

Kimi Code plans

The Kimi Code documentation lists K3.1 Pro (model ID k3.1-pro) and K3.1 Pro Highspeed (k3.1-pro-highspeed) among six models offered in the coding product. It describes the Highspeed version as having "identical model capabilities" at roughly two to three times the speed, while consuming about twice the quota of the regular version.^[7] Under Kimi Code's current membership plans, k3.1-pro requires the Plus tier or above, with the 1-million-token context reserved for Pro and above, and k3.1-pro-highspeed requires Pro or above. In Kimi Code both models accept image and video input and default to high reasoning effort.^[7]

Relation to Kimi K3

Moonshot presents K3.1 Pro as a further post-trained version of K3 rather than a new pretrained base model.^[1] The specifications Moonshot gave at launch match those published for K3: the K3 model card lists 2.8 trillion total parameters, 104 billion activated parameters, 93 layers, and a 1-million-token context window, and the K3.1 Pro announcement gives 2.8 trillion total and 104 billion active parameters with a 1-million-token context.^[2][13] The K3 model card describes that model as built on Kimi Delta Attention, a form of linear attention, and Attention Residuals; its architecture is covered in the Kimi K3 article.^[13] As of October 25, 2026, Moonshot's technical material on K3.1 Pro consisted of the tech blog, which its Kimi Code documentation calls the "technical report," and its product documentation. No K3.1 Pro model card had been published on Hugging Face.^[1][7][12]

Sources: Moonshot announcement and API guide for K3.1 Pro;^[2][4] Kimi K3 model card.^[13]

API behavior

The API guide states that K3.1 Pro "always has thinking mode enabled." Reasoning depth is set with a top-level reasoning_effort field that accepts low, high or max, with max as the API default. Streaming responses separate reasoning_content from the final content, and in multi-turn conversations and tool loops clients must send back the complete assistant message, not only its visible content.^[4] Sampling settings are fixed: temperature 1.0, top-p 0.95, n of 1, and zero presence and frequency penalties, so the guide says to omit them from requests. max_completion_tokens defaults to 131,072 and can be set up to 1,048,576.^[4]

The guide documents image and video input (images must be sent as base64 or uploaded files rather than public URLs), strict JSON-schema structured output, partial mode, tool calling with tool_choice, and dynamic tool loading through system messages.^[4] It also notes that Moonshot's built-in web search tool "is being updated and is not recommended for production workflows in the near term."^[4]

Post-training

Moonshot attributes K3.1 Pro's gains to scaling reinforcement learning during post-training. The blog says that "to enable large-scale joint RL across hundreds of downstream domains," the company optimized "the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks," so that "each increment of RL compute delivers stable, predictable returns."^[1] Moonshot says the result was "consistent gains in every domain." It has not published a training report, data mixture, or compute figures that would allow these claims to be checked independently.^[1]

Claimed capabilities

Coding and game development

Moonshot says K3.1 Pro covers coding work from small fixes in existing codebases to full-project builds. It says the model works with the Unity, Unreal Engine and Godot game engines, combining them with web search and 3D generation, and that its front-end pages, Three.js scenes and 3D assets show "stronger composition, typography, animation, and polish." The blog presents a set of game and web examples that it says K3.1 Pro created end to end.^[1]

As an example of long-horizon work, Moonshot describes an internal test in which K3.1 Pro built a human walking-dynamics simulator, with training and evaluation pipelines, on top of human-atlas, which the blog describes as an open-source anatomical model with 296 skeletal components and 402 muscle components. According to the blog, over 167 missions and 420 independent reviews the model refined a walking controller until the simulated human went from "collapsing within seconds to walking continuously with alternating strides for 11.2 seconds." When reward shaping did not improve arm motion, the blog says, the model ran open-loop probes of the simulator's dynamics, found and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller.^[1]

3D modeling and design

Moonshot reports a score of 92.7% with tools on BenchCAD, a benchmark that tests 3D reconstruction from multi-view renderings through CAD code generation, compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra in its table.^[1] The company says the model is particularly suited to parametric modeling of mechanical parts, architectural models, configurable products and furniture, and that it can write and run scripts for Blender and CAD software "without MCP connectors."^[1] With the launch, Moonshot made a feature called 3D Make available on kimi.ai and the Kimi app, which it says lets users preview and edit 3D models online, download them for 3D printing, or export them as videos or Live Photos.^[1]

Finance

Moonshot reports a score of 64.4 on OfficeQA Pro v2, a test of fiscal data analysis over real financial documents, which it describes as state of the art, and an improvement over K3 on Frontier Finance.^[1] It also reports gains on three internal finance benchmarks, with these values given in a chart on the blog:^[1]

The blog states the gains as 7.2 points, 11.6 points, and "about 4.2×" K3's performance, respectively.^[1] The blog's main benchmark table, which has no K3 column, gives the same Business Operation score of 8.9 to DeepSeek-V4.1-Flash-1017. The blog does not say whether that match is a coincidence or a labelling error.^[1]

Factual accuracy

Moonshot says it made "targeted improvements" to factual accuracy for financial analysis and research work. On Kimi Hallu Bench Hard, an internal stress test of hallucination, the blog reports that the claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro, and that the share of responses with at least one factual error was 42.4% of K3's rate.^[1] The launch post on X rounded this to "hallucination down by 65%."^[2] The blog cautions that the test uses examples "deliberately crafted to probe situations where models are prone to errors" and that "the reported rates do not represent overall error rates in daily use."^[1]

The K3.1 Pro guide on Moonshot's API platform carried a different version of this passage when it was checked on October 25, 2026. It described the evaluation as drawing on de-identified conversations in which users had flagged factual errors in responses from earlier models, and gave a fall from 23.35% to 6.85%, with an error-containing share of 46.4% of K3's. The same page's overview still contained unfilled placeholders such as "[xx.x%]" where benchmark scores belonged, and it claimed state-of-the-art results on SpreadsheetBench v2, which the tech blog does not mention.^[4] This article uses the tech blog's figures.

Cybersecurity and the open-weight decision

The blog devotes a section to cyber capability. Moonshot says it saw vulnerability-research performance rise steadily as post-training scaled, and traced the trend to cyber-related content in the general mid-training corpus and cyber-related tasks in the RL environment mix. It says preliminary experiments on internal models showed that, without data removal at either stage, downstream cyber capability "would increase substantially." For the open-weight release, Moonshot says it "identified and removed cyber-related data from both training stages" while keeping general-purpose coding tasks.^[1]

Moonshot measured the effect on xdev-bench-hard, an in-house benchmark of exploitation tasks that mostly reproduce recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL, with mitigations enabled. Each task supplies the vulnerable source, a short bug hint and the binary, and the agent must achieve arbitrary code execution in a separate sandbox and retrieve a secret string. Results are pass@3 over a random subset of runs.^[1] Moonshot's chart for this benchmark reports:^[1]

Moonshot reads the 90.5-to-81.0 gap as evidence that data removal "meaningfully reduces cyber capability relative to what the same training pipeline would otherwise produce," and the gain over K3 as evidence that general coding and debugging skills carry over to cyber tasks even after removal. It states that K3.1 Pro's cyber capability "remains behind that of frontier proprietary models that have been publicly available for at least three months."^[1]

The blog then sets out two "open-weight principles." The first is that Moonshot aims "not to release open weights that would extend the absolute frontier of cyber capability." The second is that it intends "to progressively raise the cyber capability available in open weights," arguing that open models running on users' own infrastructure let more teams audit the code they maintain, including legacy systems, and that gradual diffusion of the capability "ultimately makes all software more secure."^[1] Independent assessments of K3's cyber capability, including a joint UK AISI and US CAISI evaluation on ExploitBench, are covered in the Kimi K3 article.

Benchmarks

All results in this section are reported by Moonshot and had not been independently reproduced as of October 25, 2026. Unless the blog states otherwise, K3.1 Pro was run at max reasoning effort with temperature 1.0 and top-p 0.95.^[1] Moonshot ran K3.1 Pro on most coding benchmarks with the Claude Code harness, and on some agentic tasks with Kimi Code or Moonshot's own agent harness. Comparison scores come from a mix of sources: competitors' own technical reports and system cards, public leaderboards, Artificial Analysis, and Moonshot's own reruns. For Terminal-Bench 4.0, the blog says it took the higher of each rival's self-reported figure and its leaderboard score, except for DeepSeek-V4.1-Flash-1017, whose figure comes from DeepSeek's release blog.^[1] The proprietary columns in Moonshot's table are headed "max, with fallback," and the footnotes say "proprietary model testing may involve some fallback behavior."^[1] The scores below are copied from Moonshot's table. "n/a" marks a cell the table leaves empty.

Coding

In the blog's text Moonshot gives K3's Kimi Code Bench 3.0 Long-horizon score as 55.95%.^[1] Several coding rows were run under conditions that differ from the official setups. Moonshot ran K3.1 Pro and DeepSeek-V4.1-Flash-1017 on PostTrainBench v1.2 internally, on NVIDIA H20 GPUs rather than H100s and with DeepSeek-V4.1-Flash-1017 as the judge model; the other scores in that row come from the official leaderboard. It reports ProgramBench as a binary score in which a task counts only if its ProgramBench score is at least 95%. For Terminal-Bench-Science it substituted an older DeepSeek-V4.1-Flash score of 15.7 after its own run of the 1017 version gave 9.3.^[1]

Agentic tasks

The footnotes record several evaluation choices. Moonshot ran BrowseComp with context compaction triggered at 200,000 tokens and a 500-step cap, after which the agent was forced to answer. On Agents' Last Exam and OfficeQA Pro it scored as zero any run in which a model retrieved answers online instead of doing the required analysis; it says it saw this in the trajectories of "certain external models" on OfficeQA Pro. It took the GDPval-AA, AA-Briefcase and AutomationBench-AA scores from Artificial Analysis and the APEX-Agents scores from Mercor's leaderboard. It ran OSWorld V2.1 with its own Kimi Agent harness in the official repository.^[1]

Reasoning, knowledge and multimodal

Moonshot reports Humanity's Last Exam accuracy on the full 2,500-question set and HLE-Diamond on 1,000 questions; GLM-5.3's values, marked with an asterisk in the blog, are on the text-only subset.^[1]

Cybersecurity

Moonshot ran ExploitBench with the official harness under a 300-turn budget and reports a pass@5 "max capability score" out of 16 capability flags. It ran ExploitGym with a six-hour wall-clock budget per task. The footnotes say the Claude scores on both came from Anthropic's Claude Opus 5.5 system card and name Claude Mythos 5.1 as one of the models cited, although the table's column is headed Claude Fable 5.1.^[1]

Cost on Terminal-Bench 4.0

The blog also plots Terminal-Bench 4.0 score against average cost per task, using data it says came from tbench.ai (Artificial Analysis for DeepSeek). In that chart K3.1 Pro scores 52.8% at about US$4.16 per task at high effort and 65.7% at about US$5.25 at max effort. Claude Opus 5.5 is plotted at 64.85% for about US$14.30, GPT-6 Astra at 58.18% for about US$9.90 at max effort, and Claude Sonnet 5.5 at 61.82% for about US$22.22.^[1] The chart's Claude Opus 5.5 value differs from the 66.4% in the main table because the table uses the higher of Anthropic's figure and the leaderboard score.^[1]

Independent evaluation

As of October 25, 2026, Artificial Analysis gave Kimi K3.1 Pro a score of 55 on version 4.3.2 of its Intelligence Index, against a median of 27 for comparable models. It reported that the model generated 150 million tokens while running the index, which it called "very verbose" against a median of 87 million, and that the evaluation cost an average of US$2.11 per task. It measured output speed at 73 tokens per second, slower than the 82 average.^[8] Artificial Analysis listed the model as proprietary at that point, because the weights had not yet been released.^[8] OpenRouter's model page summarizes further third-party results from Artificial Analysis, Design Arena and Vals AI. Those include a Terminal-Bench 4.0 score of 67.2% from Artificial Analysis and 63.6% from Vals AI.^[9]

Open weights

Moonshot's blog and launch post say the full weights will be released by November 4, 2026.^[1][2] That date is a company commitment, not a completed release. When this article was checked on October 25, 2026, the public model listing for the moonshotai organization on Hugging Face did not include a K3.1 Pro repository; its most recent model was Kimi K3.^[12] Moonshot had not said which license would cover the K3.1 Pro weights. K3's weights were released under a custom Kimi K3 License with commercial conditions, so the open-weight label does not by itself say which terms will apply.^[13]

The blog links the open-weight release to its cyber-data removal: Moonshot says the removal was done "for the open-weight release of K3.1 Pro."^[1] The Kimi K3 article covers how that model's hosted launch also preceded its weight release.

Reception

SiliconANGLE reported that K3.1 Pro "can perform some tasks better than Claude Opus 5.5 and GPT-6 Astra," citing its OfficeQA Pro v2 result, on which it beat both, and its Terminal-Bench 4.0 score, which put it ahead of Astra but slightly behind Opus 5.5, and noted that its performance on cybersecurity benchmarks "significantly trails" both models. The outlet tied that gap to Moonshot's decision to remove cybersecurity data.^[10] The Next Web stressed that "the figures are Moonshot's own and have not been checked independently," and reported the xdev-bench-hard results, including K3's 33.3% and DeepSeek V4.1-Flash's 85.7%. It also noted that open models "can't be recalled once they are downloaded."^[11]

See also

• Kimi K3
• Moonshot AI
• Kimi
• Kimi Work
• Kimi K2.7-Code
• Open weights
• Mixture of experts
• Reasoning models
• Terminal-Bench

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40Moonshot AI, "Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence," kimi.ai (October 23, 2026; page modified October 24, 2026). kimi.ai/...kimi-k3-1-pro
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Kimi.ai (@Kimi_Moonshot), "Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model," X (October 23, 2026). x.com/...2113680701770068067
3. ^Kimi Developers (@KimiDevs), post announcing Kimi K3.1 Pro, X (October 23, 2026). x.com/...2113682753975890049
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Kimi API Platform, "Kimi K3.1 Pro" (API guide), accessed October 25, 2026. platform.kimi.ai/...kimi-k3-1-pro-quickstart
5. ^1 ^2 ^3Kimi API Platform, "Model Inference Pricing Explanation," accessed October 25, 2026. platform.kimi.ai/...chat
6. ^1 ^2Kimi API Platform, "Model List," accessed October 25, 2026. platform.kimi.ai/...models
7. ^1 ^2 ^3Kimi Code Docs, "Model Configuration," accessed October 25, 2026. kimi.com/...models
8. ^1 ^2Artificial Analysis, "Kimi K3.1 Pro: Intelligence, Performance & Price Analysis," accessed October 25, 2026. artificialanalysis.ai/...kimi-k3-1-pro
9. ^1 ^2OpenRouter, "MoonshotAI: Kimi K3.1 Pro," accessed October 25, 2026. openrouter.ai/...kimi-k3.1-pro
10. ^1 ^2Maria Deutscher, "Moonshot AI launches new flagship Kim K3.1 Pro model with 3T parameters," SiliconANGLE (October 23, 2026). siliconangle.com/...1-pro-model-with-3t-parameters
11. ^Ana Maria Constantin, "Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills," The Next Web (October 23, 2026). thenextweb.com/...-kimi-k3-1-pro-open-weight-cyber
12. ^1 ^2 ^3Hugging Face Hub API, model listing for the moonshotai organization, retrieved October 25, 2026. huggingface.co/...models
13. ^1 ^2 ^3 ^4Moonshot AI, "moonshotai/Kimi-K3" model card, Hugging Face, accessed October 25, 2026. huggingface.co/...Kimi-K3