Kimi-VL-A3B-Instruct
Moonshot AI · 2025-04-10 · 16.4B parameters
Kimi-VL-A3B-Instruct is Moonshot AI's non-thinking instruction-tuned Kimi-VL checkpoint for text, image, and video understanding, separately distributed from the Kimi-VL-A3B-Thinking weights. It combines a native-resolution MoonViT encoder, an MLP projector, and a Moonlight MoE language decoder; Moonshot reports rounded 16B-total and 2.8B-decoder-plus-0.4B-vision activated figures, while the released BF16 tensors contain exactly 16,407,657,776 parameters and the configuration supports 131,072 positions.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 84.5 |
| Capability | 121.0 |
| HallusionBench (OpenVLM) | 48.4 |
| MATH-Vision | 21.4 |
| MathVista (OpenVLM) | 66.0 |
| MM-Vet (OpenVLM) | 66.1 |
| MMMU Validation (OpenVLM) | 57.8 |
| MMStar (OpenVLM) | 62.0 |
| OSWorld | 8.2 |