Skip to content
Models

Kimi K3.1 Pro: what it is, and where it runs

Kimi K3.1 Pro is Moonshot AI's open-weight, natively multimodal agentic model, published on Hugging Face in November 2026, which its maker calls its most capable 3T-class model to date: 2.8 trillion parameters in a mixture of experts, 104 billion of them active per token, a context of 1,048,576 tokens, text and images in. It reasons before it answers. On Evanesca, Verda, a Finnish company, runs it on its own GPUs in Germany: your prompts go neither to Moonshot AI nor to China.

Published 6 Nov 2026

Kimi K3.1 Pro at a glance

Moonshot AI post-trained it from Kimi K3 with large-scale reinforcement learning across hundreds of domains, and replaced K3's full-attention layers with a sparse attention built for long contexts. The figures below come from the model card, read on 6 November 2026.

SpecificationKimi K3.1 Pro
ArchitectureMixture of experts: 896 experts, 16 selected per token, 2 shared; 93 layers
Parameters2.8T (104B active)
Context1,048,576 tokens
Input and outputText and images in (MoonViT-V2 vision encoder), text out
WeightsMXFP4, with MXFP8 activations — the format Moonshot trained for
Reasoning effortlow, high or max — max by default
LicenceKimi K3.1-Pro License
ReleasedNovember 2026 (Hugging Face repository created on 2 November)

What Moonshot's table claims, and where it trails

Every Kimi K3.1 Pro score on the card uses reasoning effort max, temperature 1.0 and top-p 0.95; Moonshot labels the closed models' columns “max, w/ fallback”. The table below is an excerpt. Every score is Moonshot's, not a measurement made by Evanesca; a dash means the card gives none, and the asterisk marks a score on the text-only subset.

MOONSHOT AI'S MODEL CARD · %
BenchmarkK3.1 ProFable 5.1GPT-6 AstraOpus 5.5V4.1 Flash 1017GLM-5.3
Terminal-Bench 4.065.757.958.266.442.341.8
SWE Marathon v1.151.745.642.546.340.042.5
DeepSWE v1.175.567.474.174.277.066.9
Program Bench31.553.550.065.029.119.0
HLE-Full with tools70.065.657.267.758.862.5
BrowseComp92.1—91.590.8——
HLE-Full51.157.653.659.842.643.0*
GPQA Diamond94.293.796.193.992.291.7
MMMU-Pro84.186.386.987.780.6—
OSWorld V2.179.680.775.881.869.3—

Read the rows, not the headline. K3.1 Pro leads on SWE Marathon, BrowseComp and HLE with tools, and is within a point of the best column on Terminal-Bench 4.0. It trails clearly on Program Bench and on HLE without tools; the three closed models stay ahead on MMMU-Pro, and two of them on OSWorld. It is not on the LMArena leaderboard yet.

A leaderboard is a starting point. Run your own hard prompts and count the reasoning tokens a model spends per useful answer: on Evanesca, as anywhere, those tokens are billed as output. The benchmarks page explains what Evanesca checks on each model's route.

What Evanesca checked on its route

On 6 November 2026, the day Evanesca added it, the model was called on Verda's route with no fallback:

  • Live test: an answer with its token count, the count again at the end of a streamed answer, a tool call with valid JSON arguments, and an image read correctly.
  • Benchmarks: 25 correct answers out of 25, about 93 tokens per second on a long answer, the first token in half a second — and about 2.7 seconds before the first visible word, the reasoning coming first.
  • Security: 15 adversarial prompts out of 15 handled safely, in all five categories (prompt injection, system-prompt leaks, jailbreaks, data exfiltration, harmful requests).

Its model page shows 6 of its 6 checks passed.

Where it runs on Evanesca

On Evanesca, Kimi K3.1 Pro is run by Verda — Verda Cloud Oy, a company in Helsinki, formerly DataCrunch, whose terms are under Finnish law. Verda's documentation says the model “runs in Germany”, on Verda's own GPUs, and that “inference runs without subprocessors”. Its FAQ adds that “request context is held briefly in an in-memory inference cache and is removed within 2 hours”, that “Verda doesn't use your inputs or outputs to train any AI model”, and that “no request data is sent to the model developer or any other third party”.

Your prompts go neither to Moonshot AI nor to China. Evanesca does not serve the previous generation, Kimi K3: no European operator runs it for Evanesca. Sovereignty explains why the operator, not the model, decides which law applies.

Price, and the effort dial

KIMI K3.1 PRO ON EVANESCA · VAT INCLUDED
ModelRunsInput, per million tokensOutput, per million tokens
Kimi K3.1 ProGermany, at Verda€3.024€15.12

A quick question — 300 tokens in, 400 out — costs about €0.0070, before reasoning tokens. Reasoning tokens are billed as output wherever you run a model, and K3.1 Pro thinks at max effort by default, so the effort setting is also a cost dial: Verda accepts none (no reasoning), low, high and max. At max, a small max_tokens can go entirely to thinking and the answer comes back empty — leave max_tokens unset or give it room. Open-weight reasoning models compares what thinking costs from one model to the next.

In a conversation with several turns or with tools, Moonshot asks that the whole assistant message be sent back — its reasoning_content and tool_calls included, not just its text. On Evanesca a subscription is optional: credit from €5 — see pricing.

The Kimi K3.1-Pro License

The licence lets anyone use, copy, modify, distribute, sell, fine-tune and deploy the model, with three conditions on top:

  • A company that sells access to the model as a service — giving third parties real control over inference or fine-tuning, through an API for instance — with more than 20 million US dollars of revenue over 12 months must sign a separate agreement with Moonshot AI. Products with the model built into their features, and “mere relaying of requests to models hosted by others”, are not covered.
  • A product with more than 100 million monthly active users, or more than 20 million US dollars of monthly revenue, must display “Kimi K3.1-Pro” prominently.
  • Internal use, and use through Moonshot AI's own products or certified inference partners, are exempt from both.

Open weights are what make the European route possible: any operator may run the model, and the operator decides which law your words fall under.

Frequently asked questions

When was Kimi K3.1 Pro released?

In November 2026, on Hugging Face (moonshotai/Kimi-K3.1-Pro); the repository was created on 2 November, and Verda opened access through its inference API on 4 November.

Is Kimi K3.1 Pro the same model as Kimi K3?

No: it is post-trained from Kimi K3, at the same scale — 2.8 trillion parameters, 104 billion active — with sparse attention in place of K3's full-attention layers. Evanesca serves K3.1 Pro, run by Verda in Germany, and not K3.

Can I use Kimi K3.1 Pro on Evanesca?

Yes: Verda runs it for Evanesca on its own GPUs in Germany, and your prompts go neither to Moonshot AI nor to China.

How much does Kimi K3.1 Pro cost on Evanesca?

€3.024 per million input tokens and €15.12 per million output tokens, VAT included, with reasoning tokens billed as output. A subscription is optional: credit from €5.

Can Kimi K3.1 Pro read images?

Yes. It takes text and images and writes text; its vision encoder, MoonViT-V2, has 401 million parameters. Evanesca's live test on Verda's route included an image, read correctly.