Kimi K3: Moonshot AI's 2.8-Trillion-Parameter Open-Weights Flagship

What Kimi K3 is

Kimi K3 is Moonshot AI's flagship model, released on 16 July 2026: 2.8 trillion total parameters, published under the Kimi K3 License as open weights, and self-hostable. Moonshot describe it as a natively multimodal model built on two architectural pieces they name directly — Kimi Delta Attention and Attention Residuals — with native vision and a 1 million token context window, aimed at long-horizon coding, knowledge work and reasoning.

Pricing, and the parts of it that are unusual

Input is $3 per million tokens and output is $15. There is no long-context surcharge: the flat rate holds across the entire 1 million token window, where several competitors charge more once a request passes a threshold. Cached input is $0.30 per million on an automatic context-cache hit, against $3.00 on a miss — a factor of ten. Against that: there is no free tier, no batch discount and no fine-tuning, and Moonshot currently flag their web-search tool as being updated and not recommended for near-term use. Output runs at 37.2 tokens per second with a 56.6 second time to first token.

Benchmarks

On LMArena's text leaderboard Kimi K3 has 1488 Elo (rank 19, max effort); by category, 1540 Elo in Coding (rank 13, max effort), 1499 Elo in Math (rank 21, max effort), 1532 Elo in Expert (rank 12, max effort) and 1463 Elo in Creative Writing (rank 29, max effort). Moonshot's own published figures, which are vendor claims: GPQA Diamond at 93 and a mock AIME at 97. Independently, SimpleBench records 61 and SimpleQA Verified 51.

What Moonshot say it is for

Moonshot's own positioning is autonomous, long-running technical work rather than conversation. They cite chip design, scientific computing, kernel optimisation and compiler development as tasks K3 has been applied to with limited human supervision. They also claim the model edited its own trailer from 56 source clips, a job they estimate at one to two days for an experienced editor. Named integrations include Verdent for agent harnesses and FunPlus for code generation and automated game testing. Availability outside Moonshot's own API runs through Azure AI Foundry (via Fireworks AI) and OpenRouter, alongside self-hosting.

Architecture and API settings

Moonshot's quickstart describes Kimi Delta Attention as a hybrid linear attention mechanism and states that K3 activates 16 of 896 experts through a Stable LatentMoE framework, giving roughly 2.5 times the scaling efficiency of K2. Thinking is always on; reasoning effort is set with the reasoning_effort field to low, high or max, with max the default. max_completion_tokens defaults to 131,072 and can be raised to 1,048,576. Temperature 1.0, top_p 0.95 and the penalties are fixed and should be omitted from requests. Vision input takes base64 or uploaded-file references rather than public image URLs, and video is uploaded through the files API. Multi-turn and tool-call requests must return the complete assistant message, including reasoning content.

Caching and access

Context caching is automatic once a prompt exceeds 256 tokens, with 5-minute and 1-hour cache tiers; a cache write costs $3 per million tokens and a hit $0.30. API access to K3 unlocks after a top-up of at least $1, and cumulative top-ups set the rate-limit tier. Moonshot state the full weights are on Hugging Face under the Kimi K3 License, with a technical report to follow. K3 is available in Kimi on the web, iOS, Android and HarmonyOS, Kimi Work, Kimi Code in terminals and supported IDEs, and the API. Moonshot also report that K3 designed, optimised and verified a chip for a nano model of its own architecture over 48 hours using open-source EDA tools.

Official Resources