---
license: apache-2.0
base_model: dots-studio/dots3-note-prev
base_model_relation: quantized
library_name: vllm
pipeline_tag: image-text-to-text
tags:
- nvfp4
- moe
- dots3
- multimodal
- compressed-tensors
quantized_by: Frosty40
---

# dots3-note-prev NVFP4

![hero](assets/hero.jpg)

**167G** mixed NVFP4 of [`dots-studio/dots3-note-prev`](https://huggingface.co/dots-studio/dots3-note-prev) (280B / 16B active MoE, Apache-2.0).  
Experts → NVFP4. MLA / DSA / MTP / vision / audio stay FP8 or BF16. 

vLLM `0.27+` · `Dots3NoteForCausalLM` · `--language-model-only` if you only want text.

## Benches (upstream)

Quality numbers are from **dots-studio**, not this quant.

![reasoning](assets/bench_en1.png)

![multimodal](assets/bench_en2.png)

## Files

133 shards + vision + audio. `compressed-tensors` / `nvfp4-pack-quantized` on routed+shared experts.

```bash
hf download Frosty40/dots3-note-prev-NVFP4
```