Hosted access and downloadable weights are separate milestones
Moonshot has introduced Kimi K3.1 Pro across its hosted product family. Its research index dates the release October 24, and the announcement lists Kimi’s website and apps, Kimi Work, Kimi Code and the API as access channels.
The company also promises full model weights by November 4, 2026. Those weights are the numerical parameters needed to run a model outside the provider’s hosted service. A future release commitment should not be described as a download already completed.
Kimi Code’s current documentation says K3.1 Pro is fully rolled out there. Model Current has checked the public documentation and listings, not signed into every service or verified a completed weight release.
The faster variant has a different price
The public API page lists standard K3.1 Pro at $2 per million input tokens and $10 per million output tokens. Highspeed is listed at $4 and $20 respectively, as checked on October 26. Tokens are the units of text and other encoded content that a model processes.
Moonshot describes Highspeed as offering the same capabilities with faster output. That is a provider claim, not a speed measurement reproduced by this desk.
The price per token is only one part of the cost of a task. How much material the model reads, how much it generates and how many attempts are needed also matter. The listed rates do not establish the total cost of completing a particular job.
A large context window is not an entitlement on every plan
Kimi Code’s model table lists context up to 1,048,576 tokens, with access conditions that vary by membership tier. Readers should distinguish the model’s advertised maximum from what their account can actually use.
The same documentation says the Pro models require thinking mode; requests that disable it are rejected. It also recommends a fresh session when changing from older model families, because context and cache behavior can affect results.
These are operational details, but they matter more to a migration than a headline comparison alone. A model name, an account’s limits and the client software around it are different parts of the service.
Performance claims still need a task-specific check
The release promotes gains in coding and other multi-step work and publishes benchmark comparisons. Those remain attributed company results; this article does not establish that one model is universally better than another.
For readers assessing the change, the useful distinction is concrete: hosted access can be assessed now, local deployment depends on the promised artifacts, and a working migration depends on the documented access and execution rules. None of those milestones substitutes for checking the finished work.
Sources
- Moonshot: K3.1 Pro announcementPrimary announcement lists hosted channels, the November 4 weight-release promise, Highspeed claims and company benchmark methodology. Future artifacts and comparative results are not treated as independently verified.
- Kimi Code: model configurationCurrent primary documentation: hosted rollout, thinking-only behavior, session guidance and plan-dependent context limits. No account login, configuration change or API call made.
- Kimi API: public model and price listingPublic USD listing checked October 26: standard input/output $2/$10 and Highspeed $4/$20 per million tokens. Rates are not a measured total task cost or a purchasing recommendation.
- Moonshot: research indexPrimary index dates the K3.1 Pro entry October 24, 2026. The date is not inferred from a repost, rumor or changing homepage date.
