Kimi K3
Moonshot AIReasoning ModelOpen SourceMoonshot AI's flagship multimodal reasoning model, released July 2026 — the largest open-weight model to date at 2.8T total / 104B active parameters (MoE) with a 1M token context window and the Kimi Delta Attention architecture.
Abilities
Use Cases
Available in Tools
Availability
How to Use
Pros
- Largest open-weight model released to date — frontier-level reasoning with downloadable weights
- 1M token context window with efficient Kimi Delta Attention
- Multimodal reasoning — text and vision in one model
- Competitive API pricing at $3/$15 per MTok with $0.30 cached input
Cons
- Self-hosting is impractical for most — 2.8T parameters require a data-center-class cluster
- License is modified MIT, not pure open source — large MaaS businesses need a separate agreement
- Subject of a September 2026 dispute with Anthropic over alleged routing/distillation — reputational uncertainty
What to Use It For
Perfect For
The only open-weight model at this scale — auditability and self-hosting rights that closed frontier models lack
1M context and Moonshot heritage of long-document understanding
Good For
Strong reasoning-model performance at a mid-tier API price
Not Recommended
Open weights, but the model is far too large to run outside large clusters
Try instead: Gemma 4, Qwen 3
Deep reasoning adds latency compared with lightweight chat models
Try instead: Claude Haiku 4.5, Gemini 3.8 Flash
Do Not Use For
Reasoning LLM — understands images but cannot create them
Try instead: Seedream 4, FLUX.2 Pro