Unsloth files give Qwen3.8-27B 1.5 times the speed on 24GB cards

Unsloth released new small files of the open model Qwen3.8-27B. Unsloth reports that they operate at 1.5 times the speed of the full 16-bit files. The new files have almost the same size as the full files. They fit in 24GB of card memory. Unsloth also reports an increase of the score, but it gave no test results.

How it works

The files use NVFP4, a 4-bit number format that makes a model small. Unsloth also added an adjustment to the memory of the model for the text it reads. Unsloth reports that this gives 2 times more text length.

What is necessary

The fast files are only for NVIDIA Blackwell cards:

• RTX 50X cards
• DGX Spark
• B200 and B300 cards

Owners of cards before Blackwell can use the GGUF files, a different file type.

What is not known

Unsloth gave no numbers for the increase of the score. The files operate only in vLLM at this time. The SGLang tool does not work with them.

Insight: Cheap small models make many copies affordable

Researchers report that copies of Qwen3.8-27B that use one plan increased a code test score from 66.8% to 92.4%. Microsoft also made its Decision-1 model from the small Qwen3.5-9B. This reads as a race to make Qwen the model that other teams tune, since fast files lower the cost of many copies. For builders, a speed increase of 1.5 times can be more important than a new model.