MODELS
Kimi K2
Open-weight 1T-parameter MoE from Moonshot, tuned for agentic tool use.
Specs
- Context window
- 128,000
- Max output
- 16,384
- Modalities
- text
- Tool use
- ✓
- Vision
- —
- Streaming
- ✓
- License
- modified-mit
- Released
- 2025-07-01
Pricing
- Input / 1M
- $0.60
- Output / 1M
- $2.50
- Cached input / 1M
- $0.15
Cost estimate
Kimi K2 is Moonshot's open-weight Mixture-of-Experts model released under a modified MIT license, positioned for agentic workflows, coding, and tool calling rather than pure chat. It supports 128K context and native function calling, with text-only I/O. At $0.6/$2.5 per million tokens it undercuts most closed frontier models while staying close on coding and agent benchmarks, making it a practical choice for teams that want to self-host or fine-tune.
Editor's verdict
Pick K2 when you need an open-weight model that actually behaves well as a tool-calling agent — it's one of the few non-Western open releases competitive with DeepSeek-V3 on coding and agent tasks, and the license is genuinely permissive. Trade-offs: no vision, 16K output cap is tight for long refactors, and the 1T MoE size means self-hosting needs serious GPU budget. For pure chat or multimodal work, look elsewhere.
Reviews
No reviews yet. Be the first.
Last updated: 2026-04-29