Moonshot AI
Moonshot launches Kimi K3.1 Pro with open weights due by Nov. 4
Moonshot AI's Kimi K3.1 Pro is live in the app and API with a 1 million-token context, while the open weights are promised for early November.
Moonshot AI’s Kimi K3.1 Pro is live on Kimi.com and the Kimi API with a 1 million-token context window, but the open weights have not been released. The company says they will arrive by Nov. 4.
Moonshot announced the model Oct. 24, 2026 on its Kimi research blog. It calls K3.1 Pro the first model in its K3.1 series and its most capable 3T-class open-weight model to date. K3.1 Pro is a post-trained version of Kimi K3, which Moonshot released July 16, 2026.
Where It Runs
The model is available now in Kimi.com, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API. The API names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot says the highspeed version runs 2 to 3 times as fast. Pick the model in Kimi Code with /model. Kimi Work needs desktop app version 3.1.0 or later on Windows or Apple silicon Macs.
Kimi Business administrators have to turn the model on. Access is off by default at launch. Zero Data Retention is available to API customers.
API Settings
The API takes text, image and video input. Images go in as base64 or an uploaded file ID, not a public URL. Video goes in as uploaded files. Reasoning effort has three levels, low, high and max. The default is max. Thinking cannot be turned off, though low reduces it. Temperature is fixed at 1.0 and top_p at 0.95. The max_completion_tokens setting defaults to 131,072 and can go up to 1,048,576.
Pricing is flat pay-as-you-go with no tiers by context length. Cache writes with a 5-minute or 1-hour lifetime, cache-hit input, cache-miss input and output are billed separately. API access unlocks after a top-up of at least $1. OpenRouter listed moonshotai/kimi-k3.1-pro on Oct. 23 at $2 per million input tokens and $10 per million output tokens, according to its model list read Oct. 25.
Benchmarks From Moonshot
Every benchmark below comes from Moonshot’s blog and has not been independently verified. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On Kimi Code Bench 3.0 Long-horizon, an internal test, it scored 61.2% against 55.95% for K3.
Moonshot reports claim-level hallucination of 8.3% with tools on its Kimi Hallu Bench Hard, down from 23.4% for K3. The API quickstart describes its hallucination figures as a stress test, not typical error rates. On the cyber test xdev-bench-hard (pass@3), K3.1 Pro scored 81.0% against 33.3% for K3. Moonshot says that is below the 90.5 of a version without its cyber risk-reduction measures.
Our View
We think the setting that matters most for developers is the default. Reasoning effort starts at max, and thinking cannot be switched off. A developer who sends a request without choosing an effort level gets the heaviest setting the model offers.
That matters most for the simple calls in an app, such as short classifications or quick lookups. Those requests now carry full reasoning unless the developer sets effort to low. Sampling is fixed. Effort is the main dial a developer has. The highspeed model is the other. It is a separate API name to opt into.
A team should set effort per call type instead of relying on one default. Low for the light calls and max for the hard ones is the obvious split.
That split pays off only if low holds up on the team’s own tasks.