MXFP4 Open on Hugging Face
Parameters
31B
Weights (est.)
≈ 18.3 GB
Mac memory (est.)
24 GB+
Size is estimated from the parameter count and quantisation. Memory allows a fifth extra for context plus 2 GB for the system; long contexts need more.
Run it
With mlx-vlm on an Apple silicon Mac:
pip install mlx-vlmpython -m mlx_vlm.generate --model mlx-community/P1-VL-30B-A3B-mxfp4 --image photo.jpg --prompt "Describe this image"
Details
- Base model
- PRIME-RL/P1-VL-30B-A3B (quantized)
- Vendor
- PRIME-RL
- Task
- Vision + text
- Quantisation
- MXFP4
- Licence
- apache-2.0
- Downloads
- 16 in the last 30 days · 119 all time · 0 likes
- Published
- 2026-05-18 · updated 2026-05-18 (5mo ago)
- Tags
- qwen3_vl_moe, science, physics, vision-language, reasoning, olympiad
Other MLX builds of PRIME-RL/P1-VL-30B-A3B
| Build | Quantisation | Weights | Mac memory | Downloads |
|---|---|---|---|---|
| P1-VL-30B-A3B-4bit | 4-bit | ≈ 18.3 GB | 24 GB+ | 21 |
| P1-VL-30B-A3B-mxfp4 (this one) | MXFP4 | ≈ 18.3 GB | 24 GB+ | 16 |
| P1-VL-30B-A3B-nvfp4 | 4-bit | ≈ 18.3 GB | 24 GB+ | 15 |
| P1-VL-30B-A3B-5bit | 5-bit | ≈ 22.1 GB | 32 GB+ | 15 |
| P1-VL-30B-A3B-6bit | 6-bit | ≈ 25.9 GB | 32 GB+ | 18 |
| P1-VL-30B-A3B-8bit | 8-bit | ≈ 33.5 GB | 48 GB+ | 34 |
| P1-VL-30B-A3B-mxfp8 | MXFP8 | ≈ 33.5 GB | 48 GB+ | 15 |
| P1-VL-30B-A3B-bf16 | bf16 | ≈ 62.1 GB | 96 GB+ | 38 |