Introducing Kimi K3 Kimi released Kimi K3, a 2.8-trillion-parameter open-source model with hybrid linear attention and a 1M-token context window, achieving frontier-level performance in coding, knowledge work, and reasoning. The model ranks second only to Claude Fable 5 and GPT-5.6 Sol in overall intelligence, and its full weights will be released soon. Introducing Kimi K3 Kimi K3 is Kimi’s most capable model to date, with 2.8 trillion parameters. Built on Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, it offers native visual understanding and a 1M-token context window for frontier intelligence scenarios such as software engineering, knowledge work, and deep reasoning. In our evaluations, Kimi K3 delivers frontier-level performance. Among the models tested, its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 Sol. For the complete benchmark results, see our tech blog https://www.kimi.com/blog/ . The full model weights of Kimi K3 will be released in the coming days. More details on the architecture, training, and evaluation will be published together with the Kimi K3 technical report. The 3-trillion-scale open-source model Kimi K3 is the first open-source model to reach the 2.8-trillion-parameter scale. It is the latest step in Kimi’s continued push of model-scale boundaries: in 9 of the past 12 months, Kimi models have set new records for open-source model scale. Kimi K3 is built on Kimi Delta Attention KDA and Attention Residuals AttnRes . Both architectural updates are designed to help information flow more smoothly through longer sequences and deeper models. We also further increased the sparsity of the Mixture of Experts MoE : with the Stable LatentMoE framework, the model efficiently activates 16 out of 896 experts. Together with improvements in training methodology and data recipes, these structural advances give K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively. Coding Kimi K3 is our most capable coding model to date. It sustains progress on long-horizon software engineering tasks: understanding large codebases, operating the terminal, coordinating tool calls, inspecting runtime behavior, and recovering autonomously from failed attempts with minimal human intervention. K3 improves most notably on tasks that combine software engineering, visual understanding, and spatial reasoning. It can move back and forth between source code and rendered output, using screenshots, logs, test results, and runtime state to decide the next change. This makes K3 especially well suited for game development, frontend engineering, CAD workflows, and infrastructure optimization. Knowledge work K3 pushes the boundary of end-to-end knowledge work. On the GDPval-AA v2 leaderboard, Kimi K3 scores 1687. The benchmark evaluates AI models on real-world tasks across 44 occupations and 9 major industries; Kimi K3 ranks behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and ahead of Claude Opus 4.8 Max at 1600. On AA-Briefcase, Kimi K3 scores 1527, ranking second among all models — behind only Claude Fable 5 Max and ahead of GPT-5.6 Sol Max 1495 . AA-Briefcase is a private agentic knowledge-work benchmark developed by Artificial Analysis to evaluate frontier agentic capability in long-horizon knowledge work. Thanks to the 1M context window, in a single-agent setup Kimi K3 achieves a SOTA score of 91.2 on BrowseComp without context compression or additional context-management techniques, demonstrating outstanding capability on long-horizon, high-difficulty information-seeking tasks. Get started The examples require Python 3.9+ and the OpenAI SDK. Install the SDK and initialize the client once; later Python examples reuse client . Basic call - Python - cURL Thinking effort K3 always has thinking mode enabled and supports configuring its thinking effort with the top-level reasoning effort field. Do not use the K2.x thinking parameter. Thinking effort currently supports only the max level default ; more levels are coming soon. See Thinking Effort /docs/guide/use-thinking-effort for usage.For multi-turn conversations and tool calls, add the complete assistant message returned by the API to the next request. Do not keep only content . Streaming Streaming responses provide separate reasoning content and final-answer content deltas. See Streaming Output /docs/guide/utilize-the-streaming-output-feature-of-kimi-api for details. Vision input For vision messages, content must be an array of objects, not a serialized string. See Vision Input /docs/guide/use-kimi-vision-model for formats and limits. - Local image - Video file Structured output Use json schema with strict: true to constrain the final message.content . Parse only that field, not reasoning content . Name and age schema Name and age schema Structured Output /docs/guide/response format . Partial Mode Add an assistant message with partial=True at the end of messages to continue from a text prefix. Prepend that prefix when displaying the final result. Partial Mode /docs/guide/use-partial-mode-feature-of-kimi-api . Custom tools and tool choice Use tool choice="required" on the first turn to require at least one tool call. After executing every call, return the complete assistant message and append one tool result with the matching tool call id for each call. Minimal weather agent loop Minimal weather agent loop Tool Choice /docs/guide/use-tool-choice . Dynamic tool loading Place a complete tool definition in a system message without content . The tool becomes available from that message onward. Load a calculator dynamically Load a calculator dynamically - Include the complete name , description , and parameters definition. - The declaration takes effect at its position in messages . - Keep this message in later request history; the server does not retain it. Dynamic Tool Loading /docs/guide/use-dynamic-tool-loading . 1M context and automatic caching Context caching is automatic for regular model requests; no cache ID, TTL, or extra parameter is required. Keep the long prefix unchanged so later requests can automatically attempt a cache hit. Context Caching /docs/guide/use-context-caching-feature-of-kimi-api . Official tools Official tools are integrated through Formula:- Fetch tool definitions from the Formula /tools endpoint. - Add those definitions to the Chat Completions tools field. - When the model returns tool calls , submit each function name and arguments to the Formula /fibers endpoint. - Add the complete assistant message and Fiber output as the corresponding tool message. - Call Chat Completions again until the model returns a final answer. Official Tools /docs/guide/use-official-tools for the complete client and API contract. Web search is being updated and is not recommended for use in the near term. Important limits reasoning effort currently supports only max ; K3 always has thinking mode enabled. max completion tokens defaults to 131072 and can be set up to 1048576. temperature=1.0 , top p=0.95 , n=1 , presence penalty=0 , and frequency penalty=0 are fixed; omit them from requests.- Return the complete assistant message unchanged in multi-turn conversations and tool calls. - Vision input does not support public image URLs. Use base64 or ms://