The quantized model Swift-Qwen3.8-27B-NVFP4 has been released

Developers introduced Swift-Qwen3.8-27B-NVFP4 — a specialized version of the language model, optimized for operation on graphics accelerators using mixed precision.

The model weighs 21,945 GB. The architecture implements a combination of NVFP4 and FP8 formats, which allows significantly reducing the requirements for video memory. When running on H100 accelerators, the model takes up only 20 GB of VRAM, while maintaining the original generation quality and KV-cache accuracy.

The technical solution is focused on reducing the entry threshold for deploying heavy models on server equipment. The use of NVFP4 provides higher weight packing density without critical performance loss, which is critical for systems with limited memory.

This release expands the capabilities of working with the Qwen architecture in conditions of limited computational resources. The optimization allows running the 27-billion model on configurations that previously required more memory for similar tasks.