Kimi K2

Moonshot AI’s open-weights mixture-of-experts model. 32B activated parameters, 1T training tokens. Two variants: K2-Base (for fine-tuning) and K2-Instruct (“reflex-grade” instruction-tuned for general use and agentic workflows).

Why it matters

K2 is the open-weights model that credibly claims to beat GPT-4 series and stand toe-to-toe with Sonnet 4 / Opus 4 on coding and agentic benchmarks — at an order of magnitude lower cost. Joins deepseek, qwen, llama, gemma-4 as a top-tier open-weights option.

Pricing

• $0.15 per million input tokens
• $2.50 per million output tokens

For comparison, ~10x cheaper than Sonnet at output.

Caveat: latency

At launch, hosted K2 inference is slow. Moonshot is actively optimizing. This is exactly the gap kimi-coder fills with a dedicated single-shot pipeline.

Sources

• WorldofAI Kimi Coder walkthrough — covers both K2 and the Coder app

See Also

• kimi-coder — the K2-powered app builder
• deepseek, qwen, llama, gemma-4 — sibling open-weights models
• minimax-m1 — adjacent open-weights MoE play