Nvidia released the quantized model Qwen3.8-27B-NVFP4

Company Nvidia introduced a specialized version of the language model Qwen3.8-27B-NVFP4, optimized for operation on new generation hardware. The model contains 27 billion parameters and has undergone quantization in NVFP4 and FP8 formats.

The release is focused on deployment in high-performance computing environments based on Nvidia Blackwell graphics processors. For integration and launch of the model, developers have provided full support for popular inference engines, including vLLM and SGLang. The use of these quantization formats makes it possible to significantly reduce video memory requirements while maintaining calculation accuracy.

The model is distributed under the Apache 2.0 license, which allows its use in commercial and research projects. In internal tests, the system demonstrated a result of 104 points, achieving this indicator in 2 hours of operation on the target equipment.

This release expands the ecosystem of tools for working with the Blackwell architecture, providing developers with ready-to-use weights. Optimization for 4-bit and 8-bit floating-point formats allows running large models on a smaller number of computing nodes, which reduces the cost of inference in corporate data centers.