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] Moonshot released the model weights on November 4, 2026 in its official moonshotai/Kimi-K3.1-Pro repository, under the custom Kimi K3.1-Pro License.^[24][26][33] 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] In its own testing, Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index at launch, and its Terminal-Bench 4.0 run gave the model the highest score of any model it had tested.^[14]

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. That account predates the November 4 weight release, whose model card and configuration specify the architecture directly.^[24][25][33]

Sources: Moonshot announcement and API guide for K3.1 Pro;^[2][4] Kimi K3 model card;^[13] K3.1 Pro weight-release announcement and license.^[26][33]

Released architecture

The released K3.1 Pro checkpoint retains 69 Kimi Delta Attention (KDA) layers but changes K3's 24 dense Gated MLA layers to Gated DeepSeek Sparse Attention. A lightweight indexer scores the earlier tokens, then each sparse attention layer attends to the selected 2,048 positions. The indexer still scans the history, so the selection cap does not make every part of long-context inference constant-cost.^[24][31]

Sources: released configuration and model card.^[24][25] The card describes MXFP4 weights and MXFP8 activations from quantization-aware training. Runtime documentation narrows the storage description: routed experts use MXFP4, while attention, shared experts and other components retain BF16 weights.^[24][30]

The architectural background is documented in the Kimi K3 technical report and the earlier Kimi Linear paper, which introduced KDA. Those papers evaluate their own models; their throughput and quality results are not measurements of K3.1 Pro.^[35][36]

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] On October 30, 2026, the official Kimi account posted "Introducing 3D Make in Kimi", saying the feature was available on kimi.ai and in Kimi Work and that Kimi "connects to tools like Blender and CAD tools to create models, assemble scenes, and develop games."^[21] The Kimi article describes 3D Make in more detail.

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. Some rows overlap with evaluations that Artificial Analysis and Vals AI ran themselves after the launch; those results are in the Independent evaluation section below. 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] Independent runs of HLE-Diamond gave K3.1 Pro lower scores than Moonshot's 39.4: the benchmark's maintainers at the Center for AI Safety and Scale AI reported 29.5% without tools at high reasoning effort, and Artificial Analysis listed 37.3%.^[18][19]

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

Artificial Analysis

Artificial Analysis published its results on X on October 23, 2026, the day of the launch. It ranked K3.1 Pro fourth on its Intelligence Index with a score of 55, 11 points above Kimi K3 at max effort (44), and behind only Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56), both at max effort with fallback, and Meta's Muse Wonder Full (56) at max effort.^[14] It described the gains as "large" in agentic performance and "more moderate" in other evaluations.^[14] Because the weights had not been released, Artificial Analysis listed the model as proprietary. It said K3.1 Pro would become the highest-scoring open-weights model once they are published, and that the gap between the leading open-weights and proprietary models on the index would then narrow from 12 points to 3.^[8][14] In the site's model data on October 27, 2026, the highest-scoring model labelled open weights was Xiaomi's MiMo-V2.6-Pro, at 46.^[8]

Version 4.3.2 of the Intelligence Index combines ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1.^[8] The table below gives the scores Artificial Analysis listed on October 27, 2026 for K3.1 Pro and for Kimi K3 at max effort. For the evaluations it mentioned, the launch post gave the same results rounded to whole numbers.^[8][14]

Source: Artificial Analysis model data, accessed October 27, 2026.^[8] The site marked SciCode as "under review" on that date.^[8]

The largest change from K3 was on Terminal-Bench 4.0. Artificial Analysis said K3.1 Pro achieved "the highest score of any model" on the benchmark,^[14] and in its data on October 27 the next-highest entries were Claude Sonnet 5.5 at 63.6%, then Claude Opus 5.5 and GPT-6 Astra (at xhigh effort) tied at 59.6%; Astra at max effort scored 59.1%.^[8] That order differs from Moonshot's table, which credits Opus 5.5 with 66.4% because it uses the higher of Anthropic's own figure and the leaderboard score.^[1] On GDPval-AA, K3.1 Pro ranked third, 312 Elo above K3 and behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857).^[14] The GDPval-AA, AA-Briefcase and AutomationBench-AA values in Moonshot's own benchmark table are these Artificial Analysis figures.^[1][8] Of the ten index components, AA-LCR v1.1 was the only one on which K3.1 Pro scored below K3.^[8]

In the launch post, Artificial Analysis put the cost of running the index at US$1.98 per task, "roughly equal" to Kimi K3. K3.1 Pro used 55,000 output tokens per task, 14% more than K3's 48,000, while its first-party token prices are 33% lower.^[14] The model page gave a higher figure on October 27, 2026: US$2.11 per task for K3.1 Pro and US$2.00 for K3 at max effort, a metric the site calculates from input, cache-hit, cache-write, reasoning and answer token prices.^[8] The same page said the model generated 150 million output tokens while running the index, which it called "very verbose" against a median of 81 million, and that its score was well above the median of 28 for comparable models. It measured median output speed on Moonshot's own API at 44 tokens per second and described the model as "notably slow."^[8] Tech Times, citing Artificial Analysis on October 26, reported a speed of 57.5 tokens per second.^[15] Artificial Analysis lists the model's inputs as text and image, while Moonshot's API guide also documents video input.^[4][8]

On October 27, 2026 Artificial Analysis launched version 5.0 of the Intelligence Index, which drops GDPval-AA, SciCode, Humanity's Last Exam and AA-LCR, and adds FrontierCode 2.0, Terminal-Bench-Science and HLE-Diamond, so its scores are not comparable with the v4.3.2 figures above. K3.1 Pro scored 46 on v5.0, 12 points behind the leader at launch, Claude Opus 5.5 (max) at 58. The announcement called it the top open-weights model while noting that its weights were not yet released. Of the models measured at launch, K3.1 Pro had the highest Terminal-Bench 4.0 score (67%), but it scored 37% on HLE-Diamond and 34% on Terminal-Bench-Science, against 68% for Claude Opus 5.5 (max) and 63% for GPT-6 Astra (max).^[20]

Vals AI

Vals AI scores models on its Vals Index, which weights agentic finance, coding, legal and tax tasks by each sector's share of U.S. GDP. On version 2.2 of the index, updated October 26, 2026, K3.1 Pro scored 62.52%. That put it eighth of the 28 models listed, behind Claude Sonnet 5.5 (70.85%), Claude Opus 5.5 (70.18%), Gemini 4 Argon (69.24%), Muse Wonder Full (66.79%), GPT-6 Astra (65.02%), Muse Wonder (64.88%) and GPT-6.1 Sol (63.94%), and ahead of Kimi K3 (52.95%). It ranked first in the index's open-weights view.^[16]

On Vals AI's Terminal-Bench 4.0 leaderboard, K3.1 Pro scored 63.64% with the Mini-SWE-agent harness. It placed third, behind Claude Opus 5.5 (65.15%) and Claude Sonnet 5.5 (64.14%), both also run with Mini-SWE-agent, and ahead of GPT-6 Astra (59.60%), at a cost of US$4.57 per test, against US$13.20 for Opus 5.5 and US$16.51 for Sonnet 5.5.^[17] OpenRouter's model page reproduces further Vals AI results dated October 24 and 25, 2026, including 56.0% on the Vals Finance Index, 67.0% on the Vals Legal Index, 81.9% on IOI and 84.9% on Vibe Code Bench v1.1. It also lists Elo ratings from Design Arena.^[9]

Open weights

Moonshot announced the completed weight release on November 4, 2026.^[33] The official repository includes the model card, configuration, tokenizer, license and 96 Safetensors weight shards. At revision d9e99b0792a5e93b78ea9bea62a68167e36ddd9e, inspected on November 5, the shards total 1,561,293,160,568 bytes, about 1.56 TB in decimal units. That is file storage, not an estimate of complete runtime memory, which also depends on serving configuration and cached context.^[32]

License

The Kimi K3.1-Pro License permits use, modification and redistribution subject to conditions, including retention of its notices. A licensee whose group operates a Model as a Service business and exceeds US$20 million aggregate revenue over any consecutive 12 months needs a separate agreement before commercial use. The definition excludes feature-embedded end-user products and mere forwarding to other hosts.^[26]

A commercial product or service using the software must prominently display the model name if it exceeds 100 million monthly active users or US$20 million monthly revenue. The separate-agreement and display provisions exclude internal use as defined in the license, and access through Moonshot's official products or certified inference partners. These commercial conditions accompany the released weights.^[26]

Self-hosted inference

The model card recommends vLLM, SGLang and TokenSpeed. It lists text and image modalities; hosted product documentation separately describes video input.^[24][4]

SGLang's November 4 announcement advertised up to 33,700 tokens/s per GPU on AgentX with GB300 and 450 tokens/s for a single user with DFlash2 and MegaKernel. Those are distinct aggregate and per-user measures.^[34] Its cookbook qualifies the available AgentX command snapshots: the GB300 campaign was still running, used fixed draft-acceptance lengths, and was not an InferenceX submission. The measured image also included an unmerged change; the named public image had not been checked against those commands.^[27]

vLLM's release report likewise uses workload-specific measurements. Its GB300 AgentX frontier used four speculative tokens and a synthetic acceptance length of 3.36. The report also demonstrates a 100-step LoRA reinforcement-learning run whose vLLM rollout and Megatron trainer log-probabilities matched bit for bit. These serving and training experiments do not establish general output speed or general model quality on other systems.^[29]

The vLLM recipe labels its Rubin configuration preview-only and unverified on that hardware, and says long-context accuracy of its optional NVFP4 KV cache had not been measured. Those limits apply to the recipe, even though other runtime reports describe their own Rubin tests.^[28]

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] Tech Times, writing on October 26, called the cyber-data removal "a concrete, testable safety practice" but noted that fine-tuning a released open-weight checkpoint "can in principle partially recover suppressed capabilities," so the measure's durability could be tested only once the weights are public.^[15] On October 28, 2026 (Pacific time), The Information reported, citing people with direct knowledge of the matter, that Moonshot and DeepSeek relied heavily on outputs from GPT-6 Astra when training their latest models; The Next Web, which said it had not verified the report, identified Moonshot's model as K3.1 Pro. The report is an allegation of a distillation attack.^[22][23]

Continue reading

• Kimi (chatbot)Kimi is an artificial intelligence assistant and chatbot developed by Moonshot AI.Mentioned 69 times above
• Kimi K3Kimi K3 is an open-weight multimodal reasoning model developed by Moonshot AI. Moonshot made the model available through its hosted products on July 16, 2026 and published the weights, code, configuration…Mentioned 22 times above
• TTerminal-BenchTerminal-Bench is an open benchmark for evaluating AI agents on complex, real-world tasks performed through command-line terminal interfaces.Mentioned 16 times above
• OOpen weightsOpen weights is the practice of publishing the trained parameter tensors of a neural network so that anyone can download, run, fine-tune, and redistribute the model without obtaining a separate inference…Mentioned 4 times above
• KKimi WorkKimi Work is a desktop AI agent developed by Moonshot AI for working with local files, web pages, and longer-running tasks.Mentioned 3 times above
• Moonshot AIMoonshot AI is a Beijing-based artificial intelligence company that develops the Kimi chatbot, large language models, and agent software. The company says it was founded in early 2023.Mentioned twice above

More from See also

• Kimi K2.7-Code
• Mixture of Experts (MoE)
• Reasoning models

More to read

• Large Language Model
• Artificial Analysis
• Humanity's Last Exam

Topics:Chinese AILarge Language ModelsOpen Source AIReasoning Models

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 ^37Moonshot 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 ^7Kimi.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 ^9 ^10 ^11Kimi 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 ^2Kimi Code Docs, "Model Configuration," accessed October 25, 2026. kimi.com/...models
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Artificial Analysis, "Kimi K3.1 Pro: Intelligence, Performance & Price Analysis," accessed October 27, 2026. artificialanalysis.ai/...kimi-k3-1-pro
9. ^1 ^2OpenRouter, "MoonshotAI: Kimi K3.1 Pro," accessed October 27, 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. Hugging Face Hub API, model listing for the moonshotai organization, retrieved October 27, 2026. huggingface.co/...models
13. ^1 ^2 ^3Moonshot AI, "moonshotai/Kimi-K3" model card, Hugging Face, accessed October 25, 2026. huggingface.co/...Kimi-K3
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Artificial Analysis (@ArtificialAnlys), "Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance," X (October 23, 2026). x.com/...2113688083145314416
15. ^1 ^2Earl Bensen, "Kimi K3.1 Pro Scores 4th of 215 AI Models; Moonshot Scrubs Cyberattack Data Before Open Weights," Tech Times (October 26, 2026). techtimes.com/...erattack-data-before-open-weights
16. ^Vals AI, "Vals Index" (version 2.2, updated October 26, 2026), accessed October 27, 2026. vals.ai/...vals_index
17. ^Vals AI, "Terminal-Bench 4.0" (updated October 26, 2026), accessed October 27, 2026. vals.ai/...terminal-bench-4
18. ^Center for AI Safety and Scale AI. "Introducing HLE-Diamond." Humanity's Last Exam blog. Accessed October 29, 2026. lastexam.ai/...hle-diamond
19. ^Artificial Analysis. "HLE-Diamond Benchmark." Accessed October 29, 2026. artificialanalysis.ai/...hle-diamond
20. ^Artificial Analysis. "Announcing Artificial Analysis Intelligence Index v5." October 27, 2026. artificialanalysis.ai/...sis-intelligence-index-v5
21. ^Kimi.ai (@Kimi_Moonshot), "Introducing 3D Make in Kimi," X (October 30, 2026). x.com/...2116005725814280570
22. ^Jing Yang, Juro Osawa and Qianer Liu, "OpenAI's Astra Model Becomes Top Distillation Target for China's AI Labs," The Information (October 28, 2026; paywalled, free preview). theinformation.com/...lation-target-chinas-ai-labs
23. ^Ana Maria Constantin, "DeepSeek and Moonshot trained new models on GPT-6 Astra, report says," The Next Web (October 29, 2026). thenextweb.com/...oonshot-gpt-6-astra-distillation
24. ^1 ^2 ^3 ^4 ^5 ^6Moonshot AI. "Kimi-K3.1-Pro" model card. Hugging Face. Accessed November 5, 2026. huggingface.co/...Kimi-K3.1-Pro
25. ^1 ^2Moonshot AI. Kimi-K3.1-Pro `config.json`, revision d9e99b0792a5e93b78ea9bea62a68167e36ddd9e. Accessed November 5, 2026. huggingface.co/...config.json
26. ^1 ^2 ^3 ^4Moonshot AI. "Kimi K3.1-Pro License." Accessed November 5, 2026. huggingface.co/...LICENSE
27. ^1 ^2SGLang. "Kimi-K3.1-Pro" cookbook. Accessed November 5, 2026. docs.sglang.io/...Kimi-K3.1-Pro
28. ^1 ^2vLLM. "moonshotai/Kimi-K3.1-Pro" deployment recipe. Accessed November 5, 2026. recipes.vllm.ai/...Kimi-K3.1-Pro
29. ^Inferact and the vLLM Team. "vLLM Delivers Day-0 Support for Kimi K3.1 Pro." November 4, 2026. vllm-project.github.io/...kimi-k31-pro
30. ^1 ^2AMD. "Kimi K3.1 Pro on AMD Instinct GPUs: Day-0 support and optimization." November 4, 2026. amd.com/...kimi-k3-1-pro-on-amd-instinct-gpus
31. ^1 ^2LightSeek Foundation. "Kimi K3.1 Pro at Day 0: Sparse attention on the Kimi K3 stack with TokenSpeed." Accessed November 5, 2026. lightseek.org/...tokenspeed-kimi-k3-1
32. ^Hugging Face Hub API. Official moonshotai/Kimi-K3.1-Pro repository metadata and shard sizes, revision d9e99b0792a5e93b78ea9bea62a68167e36ddd9e. Retrieved November 5, 2026. huggingface.co/...Kimi-K3.1-Pro
33. ^1 ^2 ^3 ^4Kimi.ai (@Kimi_Moonshot). Model-weight release announcement. X, November 4, 2026. x.com/...2118010528123773297
34. ^SGLang (@sgl_project). Kimi K3.1 Pro day-zero support announcement. X, November 4, 2026. x.com/...2118017145548451860
35. ^Kimi Team. "Kimi K3: Open Frontier Intelligence." arXiv:2607.24653, July 27, 2026. arxiv.org/...2607.24653
36. ^Kimi Team. "Kimi Linear: An Expressive, Efficient Attention Architecture." arXiv:2510.26692, October 30, 2025; revised November 1, 2025. arxiv.org/...2510.26692

Show all 36 references