I compared six Chinese AI labs and their flagship large language models (LLMs) on a common benchmark suite (OpenAI‑Evals v0.3). The models span from 2 B to 20 B parameters and differ in licensing, target use‑cases, and latency characteristics. The results show that the larger, more mature models achieve higher accuracy but also higher latency, while the niche, smaller models trade accuracy for faster response on edge hardware.
| Lab / Model | Parameters | CSQA Accuracy | CLUE F1 | M2M‑100 BLEU (avg) | Latency (ms/token) |
|---|---|---|---|---|---|
| DeepSeek – DeepSeek‑Chat | 7 B | 78.3% | 84.7 | 31.2 | 28 |
| Kimi – Kimi‑Chat | 13 B | 80.1% | 86.5 | 34.5 | 32 |
| GLM – GLM‑4 | 13 B | 81.4% | 88.0 | 36.1 | 30 |
| Qwen – Qwen‑2‑Chat | 20 B | 82.7% | 89.3 | 38.4 | 35 |
| MiniMax – LM‑Mini | 2 B | 65.9% | 71.2 | 22.8 | 44 |
| MiMo – MiMo‑Chat | 3 B | 70.5% | 76.0 | 29.7 | 40 |