Models
Kimi K3.1 Pro nears the top models, and Moonshot promises to open it up in November
Moonshot AI promises that its new Kimi K3.1 Pro will be available to download by November 4 at the latest. An independent measurement puts it close to the best paid models and near the top in cybersecurity.

The essentials
- Kimi K3.1 Pro can already be used on the Kimi website and through its API, the access point for programmers, but it can't be downloaded yet (MarkTechPost, Simon Willison's Weblog, Latent Space).
- Claimed by one partyMoonshot promises to release the full weights, the numbers that make up the model, by November 4, 2026 at the latest (Simon Willison's Weblog, Latent Space).
- It costs $2 per million tokens (chunks of text) of input and $10 per million of output, less than the previous K3 (MarkTechPost, Simon Willison's Weblog).
- Claimed by one partyArtificial Analysis gives it 55 points on its overall index: second place according to Simon Willison, fourth according to Latent Space (Simon Willison's Weblog, Latent Space).
Why it matters
If Moonshot keeps its word, anyone with powerful servers will be able to run, on their own, a model that comes close to the best paid ones. This suggests the gap between closed and open models is narrowing. The cybersecurity result raises a question: once the model is released, the company no longer controls what it's used for.
You can try it now on the Kimi website and app. To download it and run it yourself, you'll need business-grade hardware, not a laptop.
The details
According to MarkTechPost, Moonshot presents K3.1 Pro as a version of K3 with much more post-training. It understands text, images and video (MarkTechPost). There's a variant, Kimi K3.1 Pro Highspeed, that runs two to three times faster and costs twice as much (MarkTechPost, Simon Willison's Weblog). Simon Willison adds that the regular price is half that of Claude Opus 5.5 (Simon Willison's Weblog). It doesn't run on a normal computer: for K3, the company recommended clusters of 64 chips or more (MarkTechPost).
Almost all the performance figures come from Moonshot, and several come from internal tests. Its own charts show it falls clearly behind its rivals on scientific tasks (MarkTechPost). The company says it removed cybersecurity data from its training before opening it up (MarkTechPost). Even so, according to Latent Space, it scores 64% on Artificial Analysis's cybersecurity index, the best among publicly available models (Latent Space).
What we don't know
- Whether the weights will arrive on time and under what terms of use: the license hasn't been disclosed yet.
- Why the sources place it in different spots in the Artificial Analysis ranking, and whether anyone independent will confirm Moonshot's internal tests.
- What lies behind the accusations of undisclosed paid promotion reported by Latent Space.
Background
On October 16 we reported that Meta had released Muse Spark Open, a downloadable model that an independent measurement ranked as the best open model from the US, but behind two Chinese models.
How this story has changed
- Latent Space places the model fourth in Artificial Analysis, not second as Simon Willison said, and adds that it scores very high in cybersecurity despite Moonshot's filter.
- We published the story.
Sources
- Moonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context
- Introducing Kimi K3.1 Pro
- [AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed
- Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings
Was this story useful? Yes Not really