Kimi K3.1 Pro: benchmarks, price, API, EU hosting

What is Kimi K3.1 Pro

Kimi K3.1 Pro is the first model of Moonshot AI's K3.1 series. According to its model card, it was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and Moonshot reports gains in coding, 3D creation, financial research, factual accuracy and other agentic work. The architecture keeps Kimi K3's scale but replaces K3's full-attention layers with Gated DeepSeek Sparse Attention, which Moonshot describes as more efficient for long contexts.

The weights are shipped already quantized: MXFP4 is the format Moonshot trained for, not a compression applied afterwards, so the hosts that serve the model run the weights Moonshot published.

Benchmarks: Moonshot's own table

The scores below are an excerpt of the table on the model card. They are Moonshot's numbers, not ours: every Kimi K3.1 Pro result uses reasoning effort max, temperature 1.0 and top-p 0.95, and the proprietary columns are labelled by Moonshot « max, w/ fallback ». A dash means the card gives no score; the asterisk marks a score on the text-only subset.

Read the rows, not the headline. On this table Kimi 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 proprietary models stay ahead on MMMU-Pro, and two of them on OSWorld. The model is not yet on the LMArena leaderboard: its export published on 2026-10-28 has no K3.1 Pro entry, so our open-source benchmark page shows its GPQA Diamond score only. As on every model page, a leaderboard score is a starting point, not a verdict: run your own hard prompts.

What we measured (2026-11-06)

We wired the model on its day of arrival at Verda and checked the API with real calls rather than trusting the documentation:

• Answer, streaming and usage: a direct call returned an answer with its usage; in streaming, usage arrives on the last chunk with stream_options.include_usage, so streamed calls are metered like the others.
• Tool calling: a request with a get_weather tool returned a valid tool_call with JSON arguments.
• Images: a small red square sent as an image was read correctly (« Red »).
• Limits: Verda publishes no separate output cap. A request with max_tokens of 262,144 was accepted; one asking for 1,048,576 output tokens was refused because input plus output would exceed the 1,048,576-token context — that is the only limit we hit.

We then ran our two Frontière Verified suites through the gateway, on the Verda route with no fallback. Benchmark: passed — 25 of 25 QA answers correct, 0.50 s mean time to first token (about 2.7 s before the first visible word, after the reasoning), 93.4 tokens/s over a 4,201-token generation. Security audit: passed — 15 of 15 adversarial prompts answered safely, all five categories passed. With licence, provenance, GDPR assessment and a live API, the model carries 6 of 6 checks.

Reasoning effort and multi-turn calls

Kimi K3.1 Pro reasons before it answers, at max effort by default. Thinking tokens are real output tokens and are billed as such, so reasoning_effort is also a cost dial. The model card lists low, high and max; Verda's documentation adds none, which answers without reasoning. At max, a small max_tokens can be spent entirely on thinking — the answer then comes back empty with finish_reason: "length". Omit max_tokens, or allow plenty of headroom.

The reasoning comes back in reasoning_content, separate from the answer in content. For multi-turn conversations and tool calls, Moonshot asks that the complete assistant message — reasoning_content and tool_calls included, not just content — be passed back in messages as returned.

curl https://getfrontiereai.eu/api/v1/chat/completions \
  -H "Authorization: Bearer $FRONTIERE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "kimi-k3.1-pro", "reasoning_effort": "high",
       "messages": [{"role": "user", "content": "Review this pull request..."}]}'

Pricing and availability

Kimi K3.1 Pro is live on our catalog under the slug kimi-k3.1-pro, called with the same OpenAI-compatible request as every other model. Client price (our margin and VAT included, prepaid balance, no subscription): €3.024 per million input tokens, €15.12 per million output tokens, €0.3024 per million cached input tokens.

Our cost basis is Verda's « Inference Preview » list price for the model — 2.00 USD per million input tokens, 0.20 USD cached and 10.00 USD output — converted to euros and marked up like the rest of the catalog. The newer model is cheaper on our catalog than Kimi K3, which only runs through US infrastructure. Every price for the Kimi family, with a cost calculator, is on the Kimi API pricing page.

The Kimi K3.1-Pro License

The licence is permissive in its first clause — use, copy, modify, distribute, sublicense, sell, fine-tune and deploy — with three conditions on top:

• Model-as-a-service above 20 million USD: a licensee that offers third parties meaningful control over inference or fine-tuning (for example via an API) and whose aggregate revenue exceeds 20 million USD over any 12 months must sign a separate agreement with Moonshot AI before commercial use. The licence states that end-user products with model features embedded, and « mere relaying of requests to models hosted by others », are not model-as-a-service.
• Attribution above 100 million users or 20 million USD a month: a commercial product above either threshold must display « Kimi K3.1-Pro » prominently in its interface.
• Exemptions: internal use, and use through Moonshot AI's official products or certified inference partners, are outside both conditions.

For most companies calling the model through an API, the conditions never trigger. They matter if you plan to host the weights yourself and resell access at scale.

Sovereignty and GDPR: who runs it

On Frontière AI, Kimi K3.1 Pro is served by Verda — Verda Cloud Oy, a Finnish company in Helsinki, formerly DataCrunch — through Verda Inference, the per-token API it opened on 2026-11-04 with day-0 access to the model. Verda's documentation says the model « runs in Germany », on Verda's own GPUs, and that « inference runs without subprocessors ». The route is marked sovereign, every response carries x-sovereign: true, and keys restricted to sovereign providers can call it.

On data, Verda's FAQ states that it stores no request: context is held briefly in an in-memory inference cache and removed within 2 hours — the prefix cache behind the cached-input price — inputs and outputs are not used to train any model, and no request data is sent to Moonshot AI or any other third party. Frontière AI itself stores neither prompts nor completions, on any route. The model's sovereignty score is 90/100; our methodology explains each factor, and the GDPR vs CLOUD Act article why the operator — not the model — decides jurisdiction.

The weights come from a Chinese lab; that does not decide where your prompts go. Open weights let a European company run the model on European hardware, and the operator's jurisdiction is what applies to your data. That is the practical difference with Kimi K3 on our catalog, which runs through US infrastructure and is labelled fast access.

FAQ