Kimi K3 is live — Moonshot AI's 2.8T-parameter flagship with 1M token context, native vision and max-effort thinking, available now via kimi.ai and the Kimi API.
02 — Technical Specifications
Architecture
MoE + KDA Hybrid Attention
Sparse MoE built on Kimi Delta Attention (KDA), a hybrid linear-attention mechanism with attention residuals
Parameters
2.8T Total / 896 Experts
2.8 trillion total parameters across 896 experts, with 16 experts activated per token
Context Length
1M Tokens
Up to 1M tokens for Allegretto tier and above; Moderato tier gets 256K
Modalities
Text + Vision
Native visual understanding built in from the ground up
Reasoning
Thinking Effort: max
Ships with max thinking effort (reasoning_effort: max); low and high tiers roll out later
Code Capability
Flagship Coding
Moonshot's strongest model for coding, game/3D and knowledge tasks — coding scores surpass Claude Fable 5
API Price (per 1M tokens)
$3 In / $15 Out
Cached input just $0.30/1M — Mooncake serving keeps coding cache rates above 90%, cutting real input cost ~4×
Open Source
Open Weights · Modified MIT
Full model weights land by July 27, 2026 under a Modified MIT license — the first open 3T-class model
* All specifications are sourced from Moonshot AI's official documentation and launch announcements (July 2026).
03 — Latest News
Moonshot AI launched Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter MoE model (896 experts, 16 active) built on KDA hybrid linear attention, with native vision and up to 1M token context. API pricing lands at $3/$15 per 1M tokens, and open weights under a Modified MIT license follow by July 27.
Read MoreMoonshot AI officially teased the upcoming Kimi K3 model, promising significant improvements in reasoning, context length, and multimodal capabilities over its predecessor K2.
Read MoreEarly benchmark leaks suggest Kimi K3 achieves competitive performance against leading frontier models, with particularly strong results in mathematical reasoning and code generation tasks.
Read MoreSources close to Moonshot AI indicate that Kimi K3 will support context windows up to 1 million tokens, enabling processing of entire codebases, books, and complex multi-step workflows in a single session.
Read MoreAnalysis of Moonshot AI's strategic positioning with K3 suggests it could become a major challenger to Western AI labs, with strong performance across Asian languages and competitive English capabilities.
Read MoreAn in-depth look at Moonshot AI's journey from its founding to becoming a billion-dollar company, and how Kimi K3 represents their most ambitious technical achievement to date.
Read More04 — Frequently Asked Questions
Not affiliated with Moonshot AI.
Kimi K3 is Moonshot AI's flagship large language model, released on July 16, 2026. It's a 2.8-trillion-parameter Mixture-of-Experts model (896 experts, 16 activated per token) built on Kimi Delta Attention (KDA) hybrid linear attention, with native vision understanding and up to a 1 million token context window. Moonshot positions it as its strongest model for coding, game/3D and knowledge tasks.