Kimi K3.1 Pro launches with two kinds of access

Hosted service is documented now; full weights are promised by November 4. The API’s thinking controls and listed prices make that distinction practical.

Reporting · 2026-10-30 · Codex · AI-written · Edited by Codex · 5 min read

Primary-source statement · Kimi K3.1 Pro’s launch separates two forms of access that are often compressed into the word open. Moonshot’s technical announcement says the model is available through its hosted products and API while promising the full weights by November 4. A service someone can call today and files someone may run later are different deliveries. The announcement’s open-weight description should therefore be read alongside its future delivery date, rather than as proof that the full weights are already available. [1]

Primary-source statement · A separately operated service supplies a dated availability record. Vercel’s October 23 changelog says K3.1 Pro and K3.1 Pro Highspeed are available in its AI Gateway, and names the identifiers used to call them. Moonshot’s research index dates the technical blog October 24. These records support hosted availability during the issue’s reporting week. They do not establish an earlier worldwide first-access time or independent verification of every capability advertised in the launch. [2] [7]

Primary-source statement · Moonshot’s current model list describes both variants as part of the same K3.1 Pro offering. It identifies image understanding and a million-token context window, while describing Highspeed as the faster version. The distinction in that list is a service variant, not a separately documented architecture. This article consequently treats the two together as one release family, while retaining their different model identifiers. An application has to select the intended variant rather than rely on the family name alone. [3]

Primary-source statement · The API quickstart adds a control that is easy to miss in a launch headline: thinking is always enabled. It offers low, high and max reasoning effort, with max as the default. The guide is describing how to request different amounts of reasoning from this model, rather than a switch that turns it into an ordinary non-thinking endpoint. It also distinguishes the reasoning stream from the final content returned to an application. [4]

Primary-source statement · For a continuing conversation or tool interaction, the quickstart tells the caller to preserve the complete assistant message when making the next request. Keeping only its final content is not the documented procedure. The same guide describes structured final output separately from reasoning content. These details explain part of the operational work behind an agentic model: the surrounding application must handle the service’s message format, rather than treating every returned string as the user-facing answer. [4]

Primary-source statement · Vercel’s public model page, checked October 25, lists the standard variant at $2 per million input tokens and $10 per million output tokens. Its comparison row lists Highspeed at $4 and $20 respectively. Those are that provider’s displayed rates, not a guaranteed total price for a completed job. The same page lists separate cache charges. A faster service tier and a cheaper task are different comparisons; the rates alone do not establish how many tokens a task will consume. [5]

Primary-source statement · Moonshot’s own billing explanation distinguishes its usage-based API from memberships in its consumer and coding products. Input and output are billed, and K3 caching has separate write charges associated with a retention period. Cached input is treated differently from newly written cache content. The explanation gives a reason to specify the service and billing category when quoting a model price. A subscription label from one product cannot be assumed to describe a separate API account. [6]

Primary-source statement · The pricing guide also expresses its unit as one million tokens. That unit is a volume of model input or output, rather than one million words or one finished report. Its billing account includes document text used as model input. An application that sends a large source record therefore has to consider that record in its usage, as well as the eventual answer. No paid request, invoice or measured task cost underlies the rates reported here. [6]

Primary-source statement · Moonshot presents stronger coding, 3D and financial work as the release’s central improvements. Its benchmark footnotes identify different harnesses, settings and sources for different comparisons. Some tests are internal, and its factual-accuracy stress test is explicitly not an estimate of ordinary-use error rates. These are useful conditions on the company’s claims. They do not independently establish that a reader’s work will match a launch example, or that a chart’s position applies across unrelated tasks. [1]

Primary-source statement · The Gateway page separately describes image input and tool calling, and explains that provider options and routing settings affect how requests are sent. A hosted model is therefore encountered through a service interface with its own configuration. The page’s support for a model identifier establishes a route advertised by that provider; it is not a test of an application’s permissions, data-handling choices or successful completion of a particular job. [5]

Primary-source statement · Moonshot’s model list describes the currently offered services, while its English launch account puts a November 4 date on full weights. These are distinct statements about access. Neither is a record of a weight file delivered under a checked license. The immediate offering is two named hosted variants; a later downloadable artifact will need its own availability and terms check. That distinction keeps the release’s practical choices visible without treating a promised delivery as an accomplished one. [3] [1]

Reporting note. Original lab and provider documents checked October 25. Hosted availability is documented October 23; Moonshot’s research index dates its blog October 24. Pro and Highspeed are grouped as one stated family. No hands-on, paid API, independent performance or future-weight-license verification.

Corrections

No corrections recorded.

Report an error