Curiosity · AI Model
Kimi K2
Kimi K2 is Moonshot AI's 2025 open-weight Mixture-of-Experts model, estimated at around one trillion parameters with 32B active. It is known for ultra-long context (up to 1M tokens in some deployments), strong Chinese and English performance, and a focus on agentic coding — Moonshot markets it as one of the best open models for multi-step tool use.
Model specs
- Vendor
- Moonshot AI
- Family
- Kimi
- Released
- 2025-07
- Context window
- 1,000,000 tokens
- Modalities
- text, code
- Input price
- $0.6/M tok
- Output price
- $2.5/M tok
- Pricing as of
- 2026-04-20
Strengths
- Top-of-class open-weight SWE-bench scores at release
- Ultra-long context (up to 1M tokens)
- Aggressive pricing on the Moonshot API and via DeepInfra/Groq
- Apache-2.0-compatible licence terms
Limitations
- 1T-parameter MoE is expensive to self-host (needs 8+ H100s)
- Tooling outside China is still maturing
- Vision modality not supported in K2
Use cases
- Long-document summarisation at Chinese-market pricing
- Agentic coding pipelines on open infrastructure
- Self-hosted chat for enterprises with data-residency needs
- Research on MoE open-weight systems
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| SWE-bench Verified | ~65% | 2026-04 |
| MMLU-Pro | ~75% | 2026-04 |
| LiveCodeBench | ~55% | 2026-04 |
Frequently asked questions
What is Kimi K2?
Kimi K2 is Moonshot AI's flagship open-weight Mixture-of-Experts language model — roughly one trillion total parameters with about 32 billion active per token — optimised for long context and agentic coding.
Is Kimi K2 open-source?
Yes. Moonshot released Kimi K2 weights on HuggingFace under a permissive licence, and inference is offered by third-party providers like DeepInfra and Groq alongside Moonshot's own API.
Sources
- Moonshot AI — Kimi K2 on HuggingFace — accessed 2026-04-20
- Moonshot AI platform — accessed 2026-04-20