| Kimi K3 | DeepSeek V4 | |
|---|---|---|
| arenaElo | 1478 | 1395 |
| humanEval | 96.5 | 91 |
| sweBench | 75.5 | 62 |
| mmlu | 94 | 91.5 |
| gpqa | 87.5 | 73 |
| aime | 92 | 90 |
| Price in/out ($/M) | $3 / $15 | $1.74 / $3.48 |
The two strongest open Chinese labs, head to head. K3 is bigger (2.8T vs V4's footprint), multimodal with native vision and video, and stronger on the published coding and agentic benchmarks. DeepSeek V4 remains cheaper per token and has a longer track record of independent verification and Western hosting. For frontier capability and multimodal agents, K3. For proven, budget-friendly text workloads, V4.