Kimi K3.1 Pro: open weights, 2.8T size, license and price

Kimi K3.1 Pro is the newest flagship model from Moonshot AI, and its weights are open. The hosted model went live first. The Hugging Face repository was created on November 2, 2026, and Moonshot's launch post had promised full weights by November 4. This explainer covers the Kimi K3.1 Pro release date, license, size, price and how its benchmark claims hold up.

What Kimi K3.1 Pro is

Moonshot calls it the first model in the K3.1 series and a post-trained version of Kimi K3, trained with much more reinforcement learning compute. It is aimed at coding, long agentic tasks, finance work and 3D creation. It takes text and images as input.

There are two API variants: kimi-k3.1-pro, and kimi-k3.1-pro-highspeed, which Moonshot says runs 2 to 3 times faster. The model is also in the Kimi apps, the Kimi Work desktop app and Kimi Code.

When Kimi K3.1 Pro was released

Moonshot's launch post says the model is "available today" but carries no date in the text I could read. Artificial Analysis lists the release as October 2026. The open weights came later: the Hugging Face repo was created on November 2, and the launch post said the full weights would be out by November 4. When we checked on November 5, the repo was public, not gated, and already showed more than 11,000 downloads.

Are the weights open, and under what license

Yes, with conditions. The model card calls it open-weight and uses the "Kimi K3.1 Pro License", which Hugging Face tags as "other" rather than a standard license like Apache 2.0 or MIT. It grants broad rights to use, modify, fine-tune and redistribute, and it adds three conditions:

• A company that runs a "Model as a Service" business (giving third parties inference or fine-tuning access, for example through an API) and earns more than $20 million in total revenue over any 12 months must sign a separate agreement with Moonshot before using the model commercially.
• A product built on the model with more than 100 million monthly active users, or more than $20 million in monthly revenue, must show "Kimi K3.1-Pro" prominently in its interface.
• Neither rule applies to internal use, or to use through Moonshot's official products or certified inference partners.

Most developers and small companies fall outside both thresholds. Large API resellers do not. This is not a license that lets anyone resell it as a hosted service at scale. Read the full text before building a business on it.

How big it is

The model card lists a mixture-of-experts design, which runs only part of its network for each token. It has 2.8 trillion parameters in total, with 104 billion active per token, 93 layers, and 896 experts with 16 chosen per token. A vision encoder of about 401 million parameters handles images. Weights are quantized to MXFP4, a 4-bit format. Moonshot's launch post describes the model only as "3T-class".

Running it yourself is a large job. The card recommends vLLM, SGLang or TokenSpeed for serving, and says quantized builds exist for llama.cpp, Ollama and LM Studio.

Context window and API price

The context window is 1,048,576 tokens, about 1 million, according to the model card and the Kimi API docs. Pricing is flat, with no extra charge for long prompts. Context caching is on by default with a 5-minute lifetime, and cache writes are billed separately for 5-minute and 1-hour lifetimes. Batch requests cost 60% of the standard price. Access to the model requires a minimum $1 top-up.

When we fetched Moonshot's pricing page on November 5, the per-token price tables did not load, so we cannot quote the official rate. Artificial Analysis lists $2.00 per million input tokens and $10.00 per million output tokens, with a 90% discount for cached input. Treat that as a third-party figure and check Moonshot's page before budgeting.

Benchmarks: what Moonshot claims and what is independent

Moonshot says that with reasoning effort set to "max", K3.1 Pro scores:

• 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
• 92.1 on BrowseComp and 79.6 on OSWorld V2.1, both agentic tests.
• 94.2 on GPQA-Diamond, a graduate-level science test.
• 51.1 on HLE-Full, which is below some competitors on the card.

Several results come from Moonshot's own benchmarks, such as Kimi Code Bench, and the card notes that some rival scores were taken from the rivals' own reports. Moonshot also says its cyber capability trails the best proprietary models.

The independent number is thinner. Artificial Analysis gives it 46 on its Intelligence Index, ranked 22nd of 68 models it compares, against a median of 37 for similar models. It also finds the model slow and wordy: about 33 tokens per second, and 113 million output tokens to finish the index, above the 72.5 million median. Our DeepSeek vs Kimi comparison puts these numbers next to DeepSeek's models.

Artificial Analysis also labels K3.1 Pro "proprietary" on its page, which conflicts with the Hugging Face release. The weights are public, but under a custom license, not an open-source one.

What to watch

Independent tests of the downloaded weights, as opposed to Moonshot's hosted API, will show whether the claims hold up. So will Moonshot's pricing page, once its tables load.