Open 3T-class flagship with million-token context and native multimodality
Built for long-horizon coding, knowledge work, and complex reasoning; thinking is always on and currently supports max effort only.
Overview
Kimi K3 is Moonshot AI's new flagship model, released on July 16, 2026. It is the first model in the 3-trillion-parameter class to commit to open weights, with 2.8 trillion parameters in total. K3 targets long-horizon software engineering, end-to-end knowledge work, and complex reasoning, scaling model width, depth, and context length together through Kimi Delta Attention, Attention Residuals, and a highly sparse Mixture-of-Experts architecture. Moonshot plans to release the full weights by July 27, 2026.
Unlike the more focused K2.6 and K2.7 Code upgrades, K3 combines native vision, a one-million-token context window, terminal and tool orchestration, and structured output in one flagship model. Moonshot also states that the overall experience still trails Claude Fable 5 and GPT-5.6 Sol, while placing K3 in the frontier group on several coding and productivity evaluations.
Key capabilities
| Dimension | Detail |
|---|---|
| Parameter count | 2.8T, activating 16 of 896 experts per token |
| Context window | 1,048,576 tokens |
| Maximum output | 131,072 tokens by default, configurable up to 1,048,576 tokens |
| Input modalities | Text, image, video |
| Output modalities | Text |
| Developer features | streaming, Tool Calls, JSON Schema, Partial Mode, dynamic tool loading, automatic context caching |
Kimi K3 always reasons, and
reasoning_effortcurrently supports onlymax. Multi-turn conversations and tool calls must return the complete assistant message unchanged, including historicalreasoning_content; do not reuse the K2.xthinkingparameter. See current pricing in the model catalog.
Architecture and one-million-token context
Highly sparse MoE with two attention architecture updates
Official figures: KDA and AttnRes improve information flow across sequence length and depth; Stable LatentMoE increases sparsity.
K3 uses Kimi Delta Attention to process long sequences, Attention Residuals to improve information retrieval across depth, and Stable LatentMoE to increase expert sparsity. Moonshot says this architecture and training recipe deliver roughly 2.5x the overall scaling efficiency of Kimi K2. Automatic context caching does not require a cache ID; keeping a long prefix unchanged lets subsequent requests attempt a cache hit.
Long-horizon coding and engineering
Long autonomous execution that ends in verifiable artifacts
Cases and figures are from the official Kimi K3 technical blog and represent showcased capability ceilings.
K3 is designed for engineering runs that continue for hours or days, not just one-shot code completion. In one official case study, the model operated autonomously for 48 hours, using open-source EDA tools to design, optimize, and verify a chip. The result fit within 4 mm² and reached more than 8,700 tokens/s in simulated decoding throughput. In another research task, K3 cross-checked more than 20 papers, evaluated over 300 equations of state, and produced more than 3,000 lines of Python in about two hours.
Public evaluations also emphasize long-horizon agents. K3 scores 67.3 on DeepSWE in the mini-SWE-agent setting, while BrowseComp reaches 90.4 with a 1M context and no context-management system. The agent harnesses are not identical across every model, so these results are better read as evidence of the task ceiling than as a strict like-for-like ranking.
Knowledge work and multimodal production
A closed loop from large-scale research to visual deliverables
Official cases span industry research and gravitational-wave analysis, emphasizing concurrent agents and multi-format outputs.
K3 can combine retrieval, terminal processing, chart generation, and artifact creation into one workflow. Moonshot's ASIC industry study went through more than 120 recursive refinements, made over 2,800 search and page-fetch calls, and processed more than 11,000 pages. In a gravitational-wave analysis, K3 coordinated over 20 concurrent subagents, analyzed 391 events, and produced scientific charts, tables, and a literature review.
Native visual understanding is also part of the engineering loop. The model can iterate on front ends, games, and CAD results from screenshots, and inspect video material for animation and video-editing tasks. API vision input currently does not accept public image URLs; use base64 data or upload the file first and reference its ms:// address.
When to use it and operational notes
- Long-horizon software engineering: large-repository understanding, performance optimization, compilers, and infrastructure development.
- Research and knowledge production: multi-source retrieval, data analysis, interactive reports, spreadsheets, and presentations.
- Visually guided development: iterate on front ends, games, CAD, and digital content using screenshot or video feedback.
- Complex agent workflows: sustained tasks that benefit from tool choice, dynamic tool loading, and a one-million-token context.
K3 is sensitive to its reasoning history and should not be introduced by switching models in the middle of an existing session. It can also be overly proactive on ambiguous tasks, so production agents should define permission boundaries clearly in the system prompt or AGENTS.md. CrossModel serves moonshot/kimi-k3 through the OpenAI-compatible /v1/chat/completions endpoint.