Kimi K3.1 Pro: API Pricing, Open Weights and Whether You Can Self-Host It

Kimi K3.1 Pro is Moonshot AI’s newest flagship model, available through its API since late October 2026. Simon Willison reported the price as $2 per million input tokens and $10 per million output tokens. Moonshot has promised the open weights “by Nov 4,” which is today, but no Kimi K3.1 Pro model card had appeared on Hugging Face when this was written, so its license and exact size are still unconfirmed.

That gap matters if you run learning technology. An open-weights model can be attractive when training content is sensitive, but a model with trillions of parameters is a serious infrastructure project. This guide separates what is confirmed from what is promised, and gives you a decision checklist and cost logic you can reuse with your own quotes.

Everything below is based on Moonshot’s published pages and the Hugging Face materials for its previous model, Kimi K3. Nothing here comes from hands-on testing. All facts are current as of 4 November 2026 and several will change within days.

What is Kimi K3.1 Pro and when did it launch?

Kimi K3.1 Pro is a large language model from Moonshot AI with a 1-million-token context window. Moonshot’s launch post describes it as its “most capable 3T-class open-weight model to date,” while saying the full weights will be released by 4 November. Willison’s write-up of the launch is dated 23 October 2026.

Moonshot’s Kimi K3.1 Pro launch post says the model is available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. It also lists a “Highspeed” variant that runs two to three times faster. The post does not publish a parameter table, a license or a release date beyond “today.”

Two details from the launch post matter to a training team. Moonshot says Zero Data Retention is available for API customers. And the post gives no instructions for self-hosting, because the weights were not yet out.

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

Simon Willison’s launch post on Kimi K3.1 Pro gives $2.00 per million input tokens and $10 per million output tokens. Moonshot’s pricing page confirms a batch discount, but its price table did not render when fetched, so treat the headline rates as reported by a named outlet.

Moonshot’s API pricing documentation says batch inference for kimi-k3.1-pro is billed at 60% of the standard price, and that caching is automatic with a default five-minute lifetime. Willison adds that Highspeed costs twice as much. The table below applies simple arithmetic to those published numbers.

Your monthly bill depends on volume and on your input-to-output mix. The next table multiplies the standard and batch rates by three example volumes. These volumes are illustrations, not measurements of any real team, and they ignore caching.

Are the Kimi K3.1 Pro open weights actually released?

Not confirmed as of 4 November 2026. Moonshot promised the full weights by this date, but a check of the moonshotai Hugging Face organization showed Kimi-K3 as its newest model and no Kimi K3.1 Pro entry. Pages can lag, so check Moonshot’s Hugging Face page yourself before planning around it.

Two other checks pointed the same way. Moonshot’s own Kimi-K3 repository does not mention K3.1 Pro, and the Kimi-K3 discussion board showed no release announcement. Treat that as “not found,” not “not released,” because a repository can go live at any hour on the deadline day.

What should you do if the repository appears?

Read three things before anything else: the license file, the model card’s parameter and precision table, and the total size of the weight files. Those three facts decide whether self-hosting is permitted for you, and what it would cost. Everything in the next two sections is a way to read them quickly.

What license will Kimi K3.1 Pro use?

Unknown until the repository is public. The only evidence is precedent: Moonshot released Kimi K3 under a custom “Kimi K3 License,” not a standard open-source license. Some commentary predicted Apache 2.0 for K3.1 Pro, but that is speculation, and a similar prediction was wrong for K3.

The Kimi K3 model card says both the code and the weights are released under the Kimi K3 License and links to the full text. From that license file, as summarized when fetched:

• A business running a “Model as a Service” offering must sign a separate agreement with Moonshot once its aggregate revenue passes $20 million over any consecutive 12 months.
• Products with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display the “Kimi K3” name prominently in the interface.
• Internal use is exempt from the Model as a Service clause, and so is use through Moonshot’s official products or certified inference partners.
• The grant permits use, modification, distribution, fine-tuning and derivative works.

For a typical corporate L&D team using the model internally, those thresholds would not bite. For an LMS vendor embedding the model in a product sold to others, they could. K3.1 Pro may inherit, loosen or tighten these terms, so have counsel read the actual file.

How big is Kimi K3.1 Pro, and what hardware would self-hosting need?

Moonshot has not published K3.1 Pro’s parameter count, beyond calling it “3T-class.” Its predecessor, Kimi K3, has 2.8 trillion total parameters and 104 billion active per token, stored at four-bit precision. Simple arithmetic puts K3’s weights near 1.4 TB, so assume K3.1 Pro is similar until the model card says otherwise.

The arithmetic uses the published Kimi K3 specifications, which list MXFP4 weights, a four-bit format. Four bits is half a byte per parameter.

• Total weights: 2.8 trillion parameters × 0.5 byte = 1.4 trillion bytes, or about 1.4 TB. Real files run larger, because the card also lists higher-precision tensors.
• Active per token: 104 billion parameters × 0.5 byte = 52 GB of weights used to generate each token.

No laptop, workstation or single GPU holds 1.4 TB. For K3, the vLLM serving recipe says “at least 8x GB300” on NVIDIA, or at least 8x MI355X or MI350X on AMD, with multiple nodes for real production traffic. The recipe states no per-GPU memory figure, so this guide sizes nothing beyond it.

Two cautions follow. These are K3 numbers, not K3.1 Pro numbers. And “at least” describes a first working deployment, not a service with uptime guarantees, redundancy and an on-call engineer.

How reliable are the Kimi K3.1 Pro benchmark claims?

Most published scores are vendor-reported, meaning Moonshot ran them under settings it chose. They show K3.1 Pro competitive with the leading models, not clearly ahead. The one independent figure Willison cites, an Artificial Analysis score of 55, puts it behind Claude Opus 5.5 at 58.

The table below lists vendor-reported results from Moonshot’s launch post. Moonshot’s footnote says the evaluations used reasoning effort set to “max,” temperature 1.0 and top-p 0.95 unless noted.

One more figure is relevant to content work. Moonshot reports a claim-level hallucination rate of 8.3% with tools on its own “Kimi Hallu Bench Hard,” down from 23.4% for K3. That is an improvement, but it is not zero, and it is a vendor test on a vendor dataset.

For how the rivals in that table compare in practice, see the GPT-6 Sol vs Luna vs Astra comparison and the Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5 guide.

When does the API beat self-hosting for a training team?

For most L&D and IT teams, the API wins on cost, speed to start and maintenance. Self-hosting becomes worth exploring only when a firm rule blocks sending content to a third party, or when sustained volume is high enough to justify a dedicated multi-GPU cluster plus the engineers who run it.

Use the checklist below with your IT and legal colleagues. Each row names the question, what points to the API, and what points to self-hosting.

The data row deserves care. Moonshot says Zero Data Retention is available for API customers, but the pages reviewed do not state where the API is hosted. If your policy depends on hosting region, ask Moonshot in writing before you decide.

How do you work out the cost break-even?

Convert the API price into a blended rate for your mix, then divide your self-hosting cost by it. The result is the monthly token volume at which the two options cost the same. Every input in the calculation comes from published prices or your own quotes, never from guesses.

Start with the blended rate. If your workload sends five input tokens for every output token, one million tokens of mixed traffic costs (5 × $2 + 1 × $10) ÷ 6 = $3.33 per million at standard rates. Then divide your total monthly self-hosting cost by $3.33.

For scale, every $10,000 of monthly self-hosting cost equals about 3,000 million, or 3 billion, tokens at that blended rate. Your number needs a real quote covering hardware or rental, power, storage, networking and engineering time. Batch pricing lowers the API side by 40%, which moves the break-even point further away.

How should an L&D team evaluate Kimi K3.1 Pro for training content?

Start with the API on non-confidential material, using a small test set built from content your subject matter experts (SMEs) have already approved. Judge accuracy against that known-good version, not against how fluent the output sounds. Only a result that clears your bar justifies pricing a self-hosting project.

A practical sequence for the evaluation:

1. Pick three or four real workflows, such as drafting quiz questions, summarizing policies, rewriting course text for plain language, or tagging content to a skills taxonomy.
2. Build a test set from approved content, with a known answer or reference for each item.
3. Run the same prompts through Kimi K3.1 Pro and your current model, and have an SME score both blind.
4. Count factual errors per item, since the vendor’s own hallucination rate is not zero.
5. Record token use per workflow, then price it with the tables above.

Evaluation gives you evidence for the decision, but it does not replace SME review in production. Any model-generated training content that touches compliance or safety still needs a qualified human sign-off. For broader context on the tooling around this, see the guide to AI tools for training content creation and the wider AI tools guide.

What should you do next?

Check Moonshot’s Hugging Face page today for a Kimi K3.1 Pro repository. If it exists, read the license first, then the model card’s parameter table and file sizes, and run the arithmetic above on the real numbers. If it does not, wait for the publication before you plan.

In parallel, run the small API evaluation on non-sensitive content, and ask your vendor and legal contacts the data-residency question in writing. That costs little at the prices above, and it tells you whether self-hosting is even worth a quote.

If you already have a GPU team and a hard residency rule, request a written cluster quote now. Divide it by the blended rate to find your break-even volume, then compare it with your measured token use.