MoonshotAI: Kimi K3 (batch)

Kimi K3 is Moonshot AI’s open-weight multimodal agentic model for long-horizon coding, knowledge work and reasoning. It combines text with native visual understanding of images and video and produces a textual answer or tool request. The manufacturer highlights extended engineering sessions, large repositories and tasks that use screenshots or visual feedback alongside code. A useful brief defines the desired artifact, the evidence to use and the condition that marks the work complete.

Architecture and long context

The model card describes a sparse Mixture of Experts with 2.8 trillion total parameters and 104 billion activated parameters. Its Stable LatentMoE design selects 16 of 896 experts per token, using Kimi Delta Attention and Attention Residuals together with gated attention. The exact context length is 1,048,576 tokens. This is room for a substantial task context, not a reason to discard source boundaries or the space needed for the result.

The published native artifact uses quantization-aware training with MXFP4 weights and MXFP8 activations. That is the manufacturer’s release format, distinct from a separately supplied quantized edition or a claim about how a particular route serves the model. Open weights and a large parameter count do not turn the active-parameter figure into the total storage or hardware requirement.

Prepare a substantial coding task

Provide the relevant repository material, the existing behavior, the requested change and the compatibility constraints. Identify the files or interface that the result must affect, the checks that determine success and any supplied error or input that exposes the problem. A useful result can be a focused code change with an explanation and evidence from the checks that actually ran, rather than a description of actions that were never performed.

For a visual development task, include a screenshot or a clear view of the intended result with the relevant code and a specific discrepancy. The maker describes workflows spanning frontend work, games and CAD that use visual feedback. Ask for the change that connects the code to the observed result, and retain the reference so the delivered artifact can be checked against it.

Images, video and knowledge-work preparation

The producer’s quickstart uses typed content objects for visual messages; content is an array, not a serialized representation of that array. Its image workflow uses encoded image data or a supported file reference rather than a public image address. The video example uploads a video file and supplies the resulting reference with a question about its content. Give a legible source and specify the observation, interval or relationship that matters. The producer’s vision workflow accepts JPEG, PNG, GIF, WebP, BMP, HEIC and HEIF images, and MP4, MPEG, MOV, AVI, FLV, MPG, WebM, WMV and 3GPP video. Animated GIF or WebP can be decoded as video. Supply SVG source as text. The maker recommends images up to 4096 × 2160 and video up to 1920 × 1080; these are resolution recommendations, not rejection thresholds. Keep the request body within the documented 100M size bound and upload large or repeatedly used media as files.

For document-heavy research, label the sources and state the questions the final report should answer. For a video explanation, identify the events or sequence to analyze rather than requesting unrestricted commentary. K3 uses visual content as evidence for reasoning and returns text or tool requests. An application can pass those requests to media tools to create separate image or video artifacts.

Thinking is always enabled

K3’s producer interface always has thinking enabled. The top-level reasoning_effort setting accepts low, high and max, with max as the default. These choices are model-specific; a non-thinking setting or an effort value from another family should not be assumed. The maker’s migration guidance distinguishes K3 from K2.x by using reasoning_effort rather than carrying over the older thinking configuration.

The producer documents separate reasoning_content and final-answer content in streamed results. For a multi-turn conversation or tool workflow, it requires the complete returned assistant message to be passed back unchanged, including reasoning_content and tool_calls. Retaining only the visible final answer loses information that the model’s preserved-thinking workflow expects on the next step.

Structured final answers and partial continuation

The quickstart documents strict JSON-schema output for the final message content. Define the fields and their meanings around the artifact the application needs, and parse that final content rather than combining it with the reasoning field. A structured result can represent extracted information from supplied material or a defined report format; the schema constrains its shape, while the source evidence determines whether its values are correct.

Partial mode can continue a supplied text prefix. The producer’s example places a partial assistant message at the end of the history and combines that prefix with the returned continuation for display. Use this when the required artifact has a meaningful fixed beginning, keeping the prefix distinct from the generated remainder.

Tools and continuity of their results

K3 supports custom function tools. The producer’s example requires a tool on the first turn, preserves the complete assistant message, executes each returned call and appends a matching tool result before requesting the final answer. A proposed call is not evidence that the operation succeeded; the actual returned result becomes the evidence for the next step.

Its dynamic-loading workflow introduces a complete tool definition at a specific point in message history. The definition needs its name, purpose and parameters, and it must remain in subsequent history because the server does not retain that declaration independently. Keep tool capabilities, supplied arguments and real observations associated so a long task can continue from the work already completed.

Output controls and repeated context

The producer documents a default completion allowance of 131,072 tokens, configurable up to 1,048,576 within the model’s overall context constraints. Its current K3 API fixes temperature at 1.0, top_p at 0.95, one result, and zero presence and frequency penalties, directing callers to omit those sampling controls. These are the exact producer profile’s settings, not a transferable list of adjustable controls.

Context caching is automatic for eligible repeated prefixes, with documented 5-minute and 1-hour TTL choices and 5 minutes as the default. Partial cache blocks remain misses. Keep reusable context stable and check actual usage rather than asserting a cache hit from repeated text alone. For the final artifact, retain the supplied sources, actual tool evidence and required output structure, and check that a long answer or engineering result meets the stated completion criteria.