THE SOURCE IN CONTEXT
Publisher checkpoint and deployment references for Kimi K3, including its preserved-thinking conversation requirements.
What is the published architecture and context?
Moonshot describes Kimi K3 as a 2.8T-parameter open-weight model with native vision and a one-million-token context, built around Kimi Delta Attention and Attention Residuals. These are architecture and capability declarations. A trillion-parameter label does not tell a reader which quantized file will load, how much KV memory remains or how many users it will serve.
Source: moonshotai · pinned document
How should a conversation preserve reasoning?
Thinking is always active in the documented usage. Keep complete assistant reasoning and tool-call messages when sending subsequent turns, and set the desired reasoning effort explicitly. Dropping the reasoning history is a prompt transformation; it is not equivalent to reproducing the publisher’s agent configuration. The minimal API example’s short output limit should not be mistaken for the benchmark budget.
Source: moonshotai · pinned document
Which sampling and harness conditions matter?
The evaluation notes use max reasoning effort and temperature 1.0. Single-step tasks use top_p 0.95, while agent tasks use 1.0. Tool-free and tool-augmented scores appear together in several cells and must be read in that order. DeepSWE also distinguishes the Kimi Code harness from the separately cited mini-SWE-agent result.
Source: moonshotai · pinned document
What does native MXFP4 mean for deployment?
The card has a dedicated native MXFP4 section; select that artifact and a compatible kernel path rather than treating every four-bit conversion as equivalent. The recipe still labels its footprint as an estimate. Preserve scale format, activation precision, engine image and parallelism, and review the custom Kimi license for the intended application.
Source: moonshotai · pinned document
What does the source specify?
These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.
| Total parameters / context | 2.8T / 1 million tokens |
|---|---|
| Evaluation reasoning / temperature | max / 1.0 |
| Single-step / agentic top_p | 0.95 / 1.0 |
AI-assisted source reading and original editorial synthesis; no independent model execution.