Overview
- Tool name: Kimi K3 (kimi-k3)
- Developer: Moonshot AI
- Official website: https://platform.moonshot.ai (international) / https://platform.moonshot.cn (China)
- Category: AI Models
- License: Kimi K3 License (open weights)
Kimi K3 is Moonshot AI's most capable flagship to date: a 1M-token context window designed for frontier scenarios such as software engineering, knowledge work and deep reasoning. Weights became downloadable on July 27, 2026 under a custom Kimi K3 License (use, modification and distribution permitted, with separate-agreement and branding requirements for certain large MaaS businesses and very large commercial products).
Official model-card benchmarks: GPQA Diamond 93.5, DeepSWE 67.5, Terminal-Bench 2.1 88.3, OmniDocBench 91.1. Moonshot's own launch comparison acknowledges K3 trails the latest Claude and GPT flagships overall - the pitch is pricing and cache economics.
The Kimi lineup: K3 (flagship), K2.7 Code (coding model at CNY 6.50/27 per million tokens) and K2.6 (general, with vision and thinking modes). All are served on the Mooncake disaggregated inference architecture, where coding workloads can exceed a 90% prompt-cache hit rate.
API pricing (official, per million tokens)
- kimi-k3 (international): $3.00 input / $15.00 output; cache write $3.00, cache-hit input only $0.30
- China platform: CNY 20.00 input / CNY 100.00 output, cache hit CNY 2.00
- K2.7 Code / K2.6 (China): CNY 6.50 / CNY 27.00
Pricing verified against Moonshot's official pricing pages platform.moonshot.ai and platform.moonshot.cn/docs/pricing (October 2026).
Who it is for
- Coding agents: K3 targets long-horizon software engineering; K2.7 Code is the dedicated low-cost coding tier
- Cache-heavy workloads: with 90%+ hit rates on Mooncake, effective input cost can fall to roughly a quarter of list price
- Deep Chinese-language tasks: long-document understanding (OmniDocBench 91.1) combined with 1M context
- Self-hosting: weights are downloadable (Kimi K3 License) for compliance-minded teams
Limitations
- $15 output is about 4x the K2 generation; long-generation workloads cost noticeably more
- The license is not MIT/Apache; very large commercial products should read the extra clauses
- Moonshot's own evaluation places K3 a tier behind Claude Fable 5 / GPT-5.6 Sol overall
Frequently asked questions
Kimi K3 vs K2?
K3 is the 1M-context flagship at $3/$15; K2.7 Code / K2.6 are the 256K-context value tiers (CNY 6.5/27). K3 is clearly stronger at software engineering and deep reasoning.
Is Kimi K3 open source?
Open weights (since July 27, 2026) under the custom Kimi K3 License - not MIT/Apache; large MaaS businesses and very large commercial products face separate-agreement requirements.
What does it cost?
International: $3.00 input / $15.00 output per million tokens, cache-hit input $0.30. China: CNY 20/100.
Is there a free tier?
The Kimi web app has free usage; the API is usage-based with no permanent free tier.
*This page is not affiliated with Moonshot AI. Pricing verified as of October 2026 (official Moonshot pricing pages); confirm current figures on the official site.*
