Kimi K3.1 Pro: Specifications, Benchmarks and the Open-Weight Plan

01 / The story

What has Moonshot actually launched?

Moonshot’s official announcement introduces Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed. It describes Pro as a post-trained version of Kimi K3, meaning further training has been applied to an existing base. October 23, 2026 launch timing is corroborated by Vercel’s changelog and Simon Willison’s dated report. Our publication and source-check date is October 25, not the launch date. [1] [5] [6]

The announcement lists Kimi.ai, Kimi Work, Kimi Code and the Kimi API as access points. Those are different products and billing arrangements, not interchangeable subscriptions. An app demonstration does not establish what a developer receives through an API. [1] [3]

The strongest news is a released service paired with a future weight-release promise. Treating both as completed events would erase a distinction that matters to developers deciding between hosted access and their own infrastructure.

02 / The story

Key takeaways

• Official Pro documentation supports text, image and video input, reasoning and tool calling.
• The K3-series context limit is 1,048,576 tokens; a large window does not guarantee accurate retrieval.
• Exact 2.8T total and 104B active figures are explicitly documented for Kimi K3, Pro’s predecessor.
• Company benchmarks and outside evaluations use different tasks and should not be merged into one winner score.
• Moonshot promises Pro weights by November 4; the Pro-specific release license remains unverified.

03 / The story

Does Kimi K3.1 Pro really have 2.8 trillion parameters?

The careful answer is about lineage. Moonshot calls Pro a “3T-class” post-trained version of Kimi K3. K3’s published model card specifies 2.8 trillion total parameters and 104 billion activated parameters. The opened Pro pages do not separately tabulate those exact architecture counts, so we do not independently certify them as newly disclosed Pro specifications. “3T-class” is a rounded category, not exactly three trillion. [1] [2]

A parameter is a learned numerical value within a model. In a Mixture-of-Experts model, a routing system selects some specialist components for each token—a small piece of text—rather than using every expert every time. Total parameters describe the full stored model; activated parameters describe the parts participating in that computation. [2]

K3’s card says 16 of 896 experts are selected per token. Multiplying that fraction by the total is not a reliable way to calculate active parameters: shared components and attention also do work. Nor does a smaller active count mean you only need to store that subset. Hardware planning must account for the whole deployed model, precision and serving arrangement. [2]

04 / The story

What does the million-token context window mean?

Official K3 pricing documentation specifies 1,048,576 tokens of context. Pro’s quickstart advertises a one-million-token window and supports text, images and video. Context is the information a model can work with in a request, not a promise that every fact will be found or remembered correctly. [3] [4]

Input, instructions, tool results and generated output must fit the applicable request limits. A million-token context window is not a separate million-token output allowance added on top of a full input. Documents and visual inputs also need appropriate processing; a token is not the same as a word or a page. [3]

For a business, the useful test is whether the model can answer questions about its actual documents with traceable evidence. Start with known answers, place relevant evidence in different parts of the input and measure omissions. These are suggested evaluation steps, not a test Qlauson performed.

05 / The story

What can the model do with coding, video and 3D work?

Pro’s official documentation describes text, image and video understanding, plus tool calls. It documents always-enabled thinking with low, high and max reasoning-effort settings; max is the default. A supported video input does not mean the API natively generates a finished video, and code that builds a 3D scene is not a native downloadable mesh output. [3]

Moonshot’s examples use scripts and external tools such as Blender, CAD software and game engines. The model may coordinate that work, but those programs supply the rendering or modeling environment. Asset licenses, dependencies and human review remain part of the workflow. [1] [3]

This is similar to the distinction in our Claude video workflow article: writing instructions or code and producing a final rendered artifact are related but separate steps. A polished example does not establish repeatability across every project.

06 / The story

Which benchmarks are company claims and which are independent?

Moonshot reports 65.7% on Terminal-Bench 4.0, 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, 92.7% on BenchCAD with tools and 64.4 on OfficeQA Pro v2. These are company-reported results with benchmark-specific evaluation arrangements. They are not Qlauson reproductions or one uniform measure of everyday performance. [1]

The announcement’s evaluation notes describe maximum reasoning effort by default, temperature 1.0 and top-p 0.95, alongside tool harnesses and task-specific conditions. Some comparison figures come from other sources. A score can be useful while still requiring a close look at how the task was run. [1]

07 / The story

What do outside tests add?

Artificial Analysis lists an Intelligence Index score of 55 and output speed of 73.5 tokens per second. These are the evaluator’s reported measurements at our check, not our tests; scores and rankings can change. Output speed is not the same as total task duration, which also includes reasoning, tools and retries. [7]

Plotly’s October 23 evaluation reports 42 correct answers from 43 analytics questions, a $13.47 total run cost and a median 2.5 minutes per question. That gives useful domain-specific evidence, but its task set and software environment limit what can be generalized. It does not establish that the model never makes factual mistakes. [8]

We also found unfinished benchmark placeholders and differing hallucination figures between the official quickstart and finalized blog. We omit those disputed percentages rather than turning a documentation inconsistency into a precise accuracy claim. Check benchmark versions and original reports before comparing vendors. [1] [3]

08 / The story

How much does the Kimi K3.1 Pro API cost?

As checked October 25, Artificial Analysis and Willison corroborate reported Pro rates of $2 per million uncached input tokens and $10 per million output tokens. Willison reports Highspeed at $4 input and $20 output, with cache-hit pricing unchanged. Reported Pro cached-input pricing is $0.20 per million tokens. These are corroborated published reports, not a fully extracted official tariff: the official price table was empty in our fetched copy. Prices may change or vary by provider and applicable billing. [6] [7]

Official billing prose confirms pay-as-you-go usage separate from Kimi membership and Kimi Code, tax-exclusive prices, automatic repeated-prefix caching and separate cache-write charges. The cache-hit rate is therefore not necessarily the entire caching bill. Exact cache-write rates were not established in our review. [4]

For an illustrative uncached request using 100,000 input tokens and 10,000 output tokens, the reported Pro rates give 0.1 × $2 + 0.01 × $10 = $0.30. This excludes taxes, tools, retries and any cache charges. It is arithmetic from dated reported rates, not a quotation or measured workflow bill. [6] [7]

The official quickstart requires a successful top-up of at least $1 to unlock this flagship model. Concurrency and request/token limits depend on account tier; we did not retrieve dependable numerical tier tables. Verify the live billing and limits pages before funding a large workload. [3] [9]

09 / The story

Are Pro weights available, and what license applies?

Moonshot promises full Pro model weights by November 4, 2026. On October 25 that is a future commitment, not a verified completed release. Vercel also repeats the future date. We did not establish a public Pro checkpoint and accompanying Pro-specific license. [1] [5]

Open weights mean access to model parameters. That label does not by itself establish an open-source training process, unrestricted commercial rights or a permissive software license. The predecessor K3 has its own published license, which should not be silently transferred to Pro. [2]

If self-hosting matters, wait for the actual checkpoint, license and runtime requirements. Then check memory needs, supported hardware, redistribution conditions and ongoing operating costs. Our local-model guide explains why access to weights and practical deployment are different decisions.

010 / The story

What should developers do next?

A sensible next step is a narrow hosted evaluation with non-sensitive examples and known expected answers. Record cost, elapsed time, correct outputs and failures. Compare it with the system you already use on the same tasks rather than relying on one public leaderboard. That is Qlauson’s evaluation recommendation, not a hands-on result.

For finance, check numerical assumptions and cited documents; do not treat a high benchmark score as authorization to trade or make payments. For coding, run tests and review dependencies, secrets and generated changes. For 3D work, inspect geometry, rendering and third-party rights.

Kimi K3.1 Pro is a verified new service with a significant context window and useful external evaluation evidence. The open-weight promise is worth following, but its delivery and license still need a separate check after release. More parameters alone do not answer whether it is the right system for your work.