Kimi K3.1 Pro: A Top Open-Weight AI Model, With Caveats Worth Knowing
Moonshot's Kimi K3.1 Pro lands with 1M context, $2/$10 pricing and open weights due 4 November. Here is what holds up and what is unverified.
Kimi K3.1 Pro, released by China's Moonshot AI on 23 October 2026, is the latest sign that open-weight AI models are closing in on closed ones. It arrived in a busy week for model launches, but its mix of price, context length and a promise of downloadable weights makes it worth a closer look, as long as you separate the vendor's claims from independent tests.
What Moonshot released
Per a Yotta Labs write-up of the launch, K3.1 Pro is a post-trained version of Kimi K3 with the same architecture: a sparse mixture-of-experts design with 2.8 trillion total parameters and 104 billion active per token. In a mixture-of-experts model, only a slice of the network "wakes up" for each word, like calling the right specialist from a large hospital rather than the whole staff. It has a 1M-token context window and accepts text and image input (one benchmark site lists video too, so check the model card). Moonshot attributes the gains to larger-scale reinforcement learning.
Pricing
Reported API pricing is $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20. A "Highspeed" variant is priced at $4 and $20 and is claimed to run two to three times faster. The previous K3 was $3 and $15, so K3.1 Pro is about a third cheaper on each rate.
Performance: vendor claims versus independent checks
Moonshot's own figures put K3.1 Pro at 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5, and ahead on some other tests. These are vendor-reported and not independently replicated, and Moonshot used different setups for competitor scores, so the rows are not like for like.
Independent data is more modest. The Artificial Analysis Intelligence Index score is 55, which the Yotta Labs post compares with 58 for Opus 5.5, and a model tracker lists it fourth overall and the highest open-weight score recorded. Vals AI's own testing gives an overall Vals Index of 62.52%, ranked 8th of 27, and 63.64% on Terminal-Bench 4.0, ranked 3rd of 55. Its finance-focused Vals index was 56.01%, ranked 18th of 31. In short: strong at coding and agent tasks, good but not leading on finance-style evaluations.
Moonshot also claims a claim-level hallucination rate of 8.3%, down from 23.4% for K3. This is a vendor number and should be treated as such.
"Open weights" is not the same as "open source"
Full weights are promised by 4 November 2026, but they are not out yet. The licence has not been announced. K3 shipped under a custom Kimi K3 License that the Yotta Labs post describes as not OSI-approved, so K3.1 Pro may carry similar terms. Reading the licence matters if you plan commercial use.
Running it yourself is a serious undertaking. Yotta Labs assumes weights of around 1.56 TB, implying a multi-node GPU cluster, but Moonshot has not published a model card confirming that. The post also says Moonshot removed cyber-related training data before the open release, which may limit vulnerability-research use. Note that Yotta Labs sells a competing AI gateway, so weigh its framing accordingly.
The wider week
The release sits among others in the same period, per a model tracker: MiniMax M3.1 Flash weights on 23 October, DeepSeek V4.1-Flash on 17 October, and Google's Gemini 4 Argon rolling into AI Mode on 19 October. The trend is clear even if individual figures vary by source: strong models are getting cheaper and more available.
What to do with this
Teams building coding agents or long-document tools can trial the API at the stated prices and run their own tests rather than trusting leaderboards. Anyone planning self-hosting should wait for the licence and weights on 4 November. Finance and compliance users should note the middling finance-index rank before committing.
Key takeaways
- Kimi K3.1 Pro (23 October 2026) has a 1M-token context and reported pricing of $2 input and $10 output per million tokens; open weights are promised by 4 November, licence unannounced.
- Vendor benchmarks look strong, but independent results are more mixed: 55 on the Artificial Analysis index and 8th of 27 on Vals.
- Self-hosting looks heavy and unconfirmed; test the API first.
Sources
- [Yotta Labs: Kimi K3.1 Pro specs, price, open weights, benchmarks](https://www.yottalabs.ai/post/kimi-k3-1-pro-specs-price-open-weights-benchmarks)
- [Vals AI: Kimi K3.1 Pro](https://www.vals.ai/models/kimikimi-k3.1-pro)
- [Local AI Zone: October 2026 AI Model Updates](https://local-ai-zone.github.io/blog/October2026AIModelUpdates.html)
- [Let's Data Science: Moonshot releases Kimi K3.1 Pro API](https://letsdatascience.com/news/moonshot-releases-kimi-k31-pro-api-open-weights-due-nov-4-786e160c) (headline only; page not retrievable)
Tags: Kimi K3.1 Pro Open-Weight Models Moonshot AI LLM Pricing AI Benchmarks