Kimi K3.1 Pro API

Kimi K3.1 Pro is the first model in Moonshot AI's new K3.1 series, and it arrives with a clear claim: the strongest open-weight model anyone can download. It keeps the frame of Kimi K3, 2.8 trillion parameters with 104 billion active per token and a one-million-token context window, and puts it through far more post-training. The result scores 65.7% on Terminal-Bench 4.0 in Moonshot's own table, ahead of GPT-6 Astra and less than a point behind Claude Opus 5.5.

Moonshot announced the model on 24 October 2026 and published the full weights on Hugging Face on 4 November, as promised, with no access gate. Artificial Analysis placed it fourth on its Intelligence Index at launch with a score of 55, the highest of any open-weight model on the board, and ahead of several closed models including GPT-6 Astra.

The release is also notable for what Moonshot took out. Before publishing the weights, the company removed cyberattack-related material from training and says it deliberately reduced the model's ability to find and exploit software vulnerabilities, and it published the before-and-after scores to show how much.

Moonshot AI and the K3.1 series

Moonshot AI is a Beijing lab founded in 2023, best known for the Kimi assistant and for a run of open mixture-of-experts models that began with Kimi K2 in July 2025. Kimi K3, announced in July 2026, took the line to 2.8 trillion parameters, added native vision and a million-token window, and became the largest open-weight model released to date.

K3.1 Pro is not a new pretraining run. Moonshot describes it as a post-trained version of K3 built with "substantially larger post-training compute", including large-scale reinforcement learning across hundreds of task domains. Moonshot also markets it as a "3T-class" model; the model card gives the exact figure of 2.8 trillion total parameters.

It launched alongside a sibling, Kimi K3.1 Pro Highspeed, which Moonshot says runs two to three times faster with identical capabilities. Both are available on kimi.ai, in the Kimi apps and on Moonshot's API, which accepts both OpenAI- and Anthropic-style requests. On Unicorn API the model id is kimi-k3-1-pro.

Open weights and the licence

The weights live in the moonshotai/Kimi-K3.1-Pro repository on Hugging Face: 96 safetensors shards, about 1.56 terabytes in total. They ship under the Kimi K3.1 Pro License, which follows the pattern Moonshot set with K3. It is a permissive, MIT-style grant to use, modify, fine-tune and redistribute the model, with two conditions aimed only at the largest players: model-as-a-service businesses above a revenue threshold need a separate agreement with Moonshot, and products with very large user bases must display the model's name. For nearly every team, it is open in every practical sense.

What changed from Kimi K3

Sparse attention in place of full attention

K3 combined 69 layers of Kimi Delta Attention, Moonshot's linear-attention mechanism, with 24 layers of Gated MLA, a compressed form of full attention. K3.1 Pro keeps the 69 linear layers and replaces the 24 full-attention layers with Gated DeepSeek Sparse Attention (Gated DSA). Sparse attention lets each token attend to a selected subset of earlier positions rather than all of them, and Moonshot made the change specifically for more efficient long-context modelling. With a million-token window, those 24 full-attention layers were where compute grew fastest as a request got longer.

Much more post-training

The bulk of the improvement comes from reinforcement learning. Moonshot reports gains across its coding suite, including 5.25 points on its internal Kimi Code Bench, and new highs on long-horizon agentic tests. On SWE-Marathon v1.1, a benchmark of extended software tasks, K3.1 Pro scores 51.7, the best figure in Moonshot's comparison, and 75.5 on DeepSWE v1.1.

Fewer false claims

K3's weak point was confident guessing: independent tests found its hallucination rate rising over K2.6. Moonshot says K3.1 Pro cuts false claims on its own factuality measure from 23.4% to 8.3%. Artificial Analysis, using a different test, still measures a 44% hallucination rate, so the improvement is real but the problem is not solved.

A deliberate cut to hacking ability

Moonshot's internal software-exploitation test tells the story. An unrestricted K3.1 checkpoint scored 90.5%; the released model scores 81.0%, after cyberattack-related material was removed from training. K3 scored 33.3% on the same test. In other words, the public model is still far more capable at security work than its predecessor, but Moonshot chose not to release the strongest version it had. On the model card it reports 74.7 on ExploitBench and 81.0 on xdev-bench-hard.

How K3.1 Pro is built

The architecture is a mixture of experts with 93 layers. Each token is routed to 16 of 896 experts, plus two shared experts, which is how a 2.8-trillion-parameter model runs with only 104 billion parameters active. Vision comes from MoonViT-V2, a 401-million-parameter encoder trained with the language model rather than bolted on afterwards, and Moonshot describes the model as understanding text, images and video.

Like K3, it was trained to run in low precision, with MXFP4 weights and MXFP8 activations. That keeps the download at 1.56 terabytes rather than several times that, but self-hosting still needs datacenter-scale hardware. Moonshot recommends vLLM, SGLang or TokenSpeed for serving. It has not published its training data, token count or knowledge cutoff.

Thinking is always on. Moonshot's API offers low, high and max reasoning effort, with max as the default and the setting used for every published benchmark.

Kimi K3.1 Pro specs

Benchmarks and first reactions

Moonshot's launch table puts K3.1 Pro close to the best closed models on agentic coding:

On vision, the model card lists 84.1 on MMMU-Pro, 89.0 on CharXiv and 95.4 on BabyVision. Independent runs broadly agree. Artificial Analysis measured 67% on Terminal-Bench 4.0, ahead of Claude Sonnet 5.5 at 64%, and ranked the model third on its GDPval-AA knowledge-work test at 1,846 Elo, behind only Claude Opus 5.5 and Muse Wonder Full.

The same tables show where it falls short. On Terminal-Bench-Science it scores 38.1 against 63.3 for Claude Opus 5.5 and 68.1 for GPT-6 Astra, and on ProgramBench 31.5 against 65.0 for Opus 5.5. Scientific computing and building whole programs from a specification are clearly weaker than everyday agentic coding.

The launch was not free of controversy. In its first days, an X user posted what were said to be campaign instructions asking creators to promote Kimi without labelling the posts as paid partnerships. The document was not authenticated and no company response had been confirmed when it circulated; it has no bearing on the model's results, but it coloured some of the early discussion.

Where Kimi K3.1 Pro is strong

Long agentic coding. Leading SWE-Marathon and sitting within a point of Opus 5.5 on Terminal-Bench 4.0 make it a serious choice for coding agents that run for many steps in a terminal or repository.

Office and financial documents. The OfficeQA Pro v2 result and the GDPval-AA placement point to strong knowledge work: spreadsheets, reports, filings and business analysis.

Research and browsing. A BrowseComp score of 92.1 and a million-token window suit deep-research agents that must read a lot before they answer.

Images and charts. Native vision with high CharXiv and MMMU-Pro scores means screenshots, charts, scanned pages and diagrams can go straight into the conversation.

3D and CAD work. BenchCAD at 92.7 with tools reflects a strength Moonshot highlights in its launch: generating and iterating on 3D and CAD geometry.

Control over deployment. The weights are public. Teams with the hardware can host the model themselves, fine-tune it and keep data in-house within the licence terms.

Where it falls short

Science and from-scratch programs. The Terminal-Bench-Science and ProgramBench gaps are large, not marginal. For scientific computing pipelines or generating a complete program from a spec, the closed leaders are clearly ahead.

Hallucination is lower, not gone. Moonshot's 8.3% and Artificial Analysis's 44% come from different tests, but together they say the same thing: ground factual answers in documents or search results.

Always thinking. There is no non-reasoning mode, and benchmarks were run at maximum effort. Simple questions still get a reasoning pass, which adds latency and output tokens.

Fragile history. Like K3, the model expects the complete assistant message, including reasoning_content and tool calls, to be passed back unchanged in multi-turn and tool loops. Dropping it degrades quality.

Security research is limited by design. The reduced exploitation ability is a deliberate safety choice. Defensive security teams should expect it to be less capable than the unrestricted checkpoint Moonshot measured.

How to prompt Kimi K3.1 Pro

Treat it as an agent that will run with a brief. State the goal, the constraints and what it must not do, put hard limits in the system prompt, and let it work. Start sessions with K3.1 Pro rather than switching to it midway, and return each assistant message as received. Keep the start of long conversations stable so cached context can be reused.

For a long coding task with guardrails:

You are migrating this service from callbacks to async/await. Do not change public function signatures or add dependencies. Run the test suite after each file, stop and ask before deleting anything, and finish with a table of every file changed and why.

For document analysis with checkable output:

Here are the four quarterly reports and the staffing spreadsheet. For each department, compare planned and actual headcount, quote the sentence that explains any gap over 10%, and list the page or cell you took each number from.

For vision:

This is a screenshot of a dashboard and a photo of the whiteboard sketch for its redesign. Write the HTML and CSS for the new layout, keeping every metric from the screenshot, and note anything in the sketch you could not read.

Ask for sources, page numbers or cell references whenever facts matter; it makes hallucinations easy to catch.

Kimi K3.1 Pro against the catalog

Claude Opus 5.5 is the model K3.1 Pro is measured against. It edges ahead on Terminal-Bench 4.0 and BenchCAD, leads GDPval-AA, and is far stronger on scientific computing and ProgramBench. Pick Opus 5.5 when the hardest tasks have to be right first time; pick K3.1 Pro for long agentic coding and document work when open weights matter.

GPT-6 Astra trails K3.1 Pro by seven points on Terminal-Bench 4.0 and sits below it on the Intelligence Index, but leads clearly on Terminal-Bench-Science. For science-heavy agents, Astra is the better test.

Kimi K3 is the direct predecessor. K3.1 Pro keeps its size and context, adds sparse attention and much more post-training, and makes fewer false claims, so there is little reason to start new work on K3.

DeepSeek V4.1 Flash is the other major Chinese open-weight option in the catalog, and one of the models K3.1 Pro pushed down the Intelligence Index. DeepSeek remains worth testing for high-volume work; K3.1 Pro is the one for the hardest open-weight tasks. K3.1 Pro is also the more restricted of the two on offensive security: DeepSeek V4.1 Flash scored 85.7% on the exploitation test where the released K3.1 Pro scored 81.0%.

Getting started

Open the Playground to try Kimi K3.1 Pro without code. To build with it, create a key on the API keys page and call the OpenAI-compatible /v1/chat/completions endpoint with model: "kimi-k3-1-pro". Send images as image_url message parts and tools in the standard tools array. The API tab on this page has full request examples, and the docs cover streaming, tool calls and errors.

Verdict

Kimi K3.1 Pro shows how much a frontier model can still gain from post-training alone. On the same 2.8-trillion-parameter base as K3, Moonshot has closed most of the gap to Claude Opus 5.5 on agentic coding, cut false claims sharply, swapped in sparse attention for more efficient long context and released the whole thing as open weights. Its weak spots are specific and measurable: scientific computing, whole-program generation, lingering hallucinations and an always-on reasoning pass. For coding agents, document-heavy work and anyone who wants frontier capability they can download and run, it is the open model to beat.