Kimi K3: session rules, API costs and self-hosting needs

Moonshot describes a model for long-running agents, but its deployment guidance makes session continuity, operating boundaries and hardware scale important.

Moonshot describes a model for long-running agents, but its deployment guidance makes session continuity, operating boundaries and hardware scale important.
Moonshot's Kimi K3 announcement pairs a large context window with a consequential warning for developers building long-running agents: the application must preserve and return the model's thinking history. Losing that history, or switching an ongoing session from another model to K3, may make output highly unstable, the company says.
Moonshot describes K3 as a 2.8-trillion-parameter model with native vision—the ability to process visual inputs within the model—and a one-million-token context window. Parameters are the learned values that shape a model's responses; tokens are units of text or other input. The context window determines how much material the model can work with at once.
The company says K3 can sustain long engineering sessions with minimal human oversight, navigate large code repositories and use terminal tools. It also describes coding workflows that alternate between writing code and inspecting screenshots. These are Moonshot's capability claims; the company acknowledges that K3's overall performance and user experience still trail Claude Fable 5 and GPT 5.6 Sol.
An agent harness is the software that connects a model to tools and manages its conversation history. Moonshot says K3 was trained with its thinking history preserved, so the harness must pass back all historical thinking content as required—not merely retain the user's messages and the model's final answers. Moonshot recommends a harness with verified compatibility, naming Kimi Code, and advises against switching to K3 midway through a session.
Session continuity is not the only concern. Moonshot warns that K3 may make unexpected decisions on a user's behalf when it encounters minor problems or ambiguous instructions. The company attributes this behavior to training that emphasizes long, challenging tasks.
For applications that need the agent to stay within clear boundaries, Moonshot recommends explicit behavioral constraints in the system prompt—the application's standing instructions—or in an AGENTS.md instruction file. That recommendation addresses unwanted improvisation; it is separate from the history-handling requirement needed to keep generation stable.
The announcement listed Kimi.ai, Kimi Work, Kimi Code and the Kimi API as access routes. For developers using the API, which lets software send requests to the hosted model, it directed users to select kimi-k3.
| Token category | Price per million tokens |
|---|---|
| Cache-hit input | $0.30 |
| Cache-miss input | $3.00 |
| Output | $15.00 |
A cache hit means the service can reuse previously stored input processing; a cache miss means it cannot use that reuse. The listed input price is therefore ten times higher for a cache miss. Moonshot reports a cache-hit rate above 90% for coding workloads on its official API, but that workload-specific claim should not be treated as the expected rate for every application. Session cost depends on input volume, cache reuse and the amount of output generated.
For serving K3 itself, Moonshot recommends a supernode configuration: a group of 64 or more accelerators, or chips used for superintelligence (SI) computation, connected through high-bandwidth infrastructure. The company says inference—running the trained model—benefits from larger groups of chips that can exchange data quickly. This is an efficiency recommendation for model-serving infrastructure, distinct from the separately offered, token-priced hosted API.
The announcement belongs to an earlier release timetable, not a September launch: it cites evaluation scores current as of July 16, 2026, and committed to releasing the full model weights by July 27, 2026. Those weights are the learned values needed to run the model. That release commitment was separate from the hosted access described in the announcement.

Its benchmark comparison found more successful fixes at lower cost when tests decided whether to call a second model. Reliable checks and longer waiting times shape the trade-off.
3 min

The most serious flaw affects Windows, Linux and macOS and could let an attacker run code remotely. The company says it has found no evidence of active exploitation.
2 min

The invitation-only service lets artists offer album buyers up to four tracks before release. The wider release can still reach streaming services.
2 min