@0xkeenz: Interesting, I was just studying the differences between Unsloth and NVIDIA's Qwen3.6 27B NVFP4 yesterday, and today Unsloth updated! The new Unsloth's quantization approach is very similar to NVIDIA's official solution: instead of choosing between BF16 and ...

X AI KOLs Timeline Tools
unsloth quantization nvfp4 fp8 qwen3.6 gpu-memory model-optimization

Summary

Unsloth releases a new version of Qwen3.6 27B NVFP4 quantization scheme, introducing FP8_E4M3 intermediate precision layer and refined weight protection, achieving 2.5x speed improvement on 24GB VRAM, while improving accuracy and tool-calling capabilities.

Interesting, I was just studying the differences between Unsloth and NVIDIA's Qwen3.6 27B NVFP4 yesterday, and today Unsloth updated! The new Unsloth's quantization approach is very similar to NVIDIA's official solution: instead of choosing between BF16 and NVFP4, it introduces FP8_E4M3 as an intermediate precision layer — sensitive weights are protected with FP8_E4M3, while less sensitive weights continue to use NVFP4. However, the weight protection in the new Unsloth seems more refined than NVIDIA's official version: in addition to the Attention and Linear-Attention main projections using FP8, the LM Head and the last 8 layers of MLP are also upgraded to FP8. So from a weight allocation perspective, this is probably the highest quality Qwen3.6 27B NVFP4 version available!
Original Article
View Cached Full Text

Cached at: 07/11/26, 12:19 AM

Interesting — just yesterday I was comparing Unsloth and NVIDIA’s Qwen3.6 27B NVFP4, and today Unsloth released an update!

The new Unsloth’s quantization approach is very similar to NVIDIA’s official solution: instead of choosing between BF16 and NVFP4, it introduces FP8_E4M3 as an intermediate precision layer — sensitive weights are protected with FP8_E4M3, while low-sensitivity weights continue to use NVFP4.

However, Unsloth’s new version seems more refined in weight protection than NVIDIA’s official version: in addition to attention and linear-attention main projections using FP8, the LM head and the last 8 MLP layers are also upgraded to FP8.

So from a weight allocation perspective, this may be the highest quality Qwen3.6 27B NVFP4 version to date!

Unsloth AI (@UnslothAI): We’re releasing new Qwen3.6 quants that run 2.5× faster on your GPU.

Qwen3.6-27B NVFP4 runs on 24GB VRAM. 35B-A3B can hit 17,561 tok/s (B200).

We also improved accuracy, tool calling, agent use, and looping.

Guide: https://t.co/EEQIlFrR0c Qwen3.6 NVFP4:

Similar Articles