ISTA-DASLab released quantized versions of Qwen3.8-Flash-Next

The research group ISTA-DASLab presented Qwen3.8-Flash-Next-GSQ-RCO-GGUF — a set of quantized weights for the Qwen3.8-Flash-Next neural network. The release includes optimized model variants prepared using GSQ-RCO nonlinear quantization methods.

The package includes quantized versions of the Q2_0, IQ2_XS, and IQ3_XXS formats. The compression level varies in the range of 2.4 to 3.0 bits per parameter (bpw), allowing the model to run on devices with limited video memory. Additionally, a mmproj file of size 0.91 GB is included in the repository to support multimodal features.

The weight preparation process took 2 hours, after which the model received a score of 129 points in local performance tests. The use of GSQ-RCO methods (Gradient-based Scalar Quantization with Rank-Constrained Optimization) is aimed at minimizing accuracy losses at extremely low bitrates.

This release expands the capabilities of local deployment of high-performance models in the Qwen series. The use of the GGUF format ensures compatibility with popular inference engines, allowing the neural network to be deployed on consumer equipment without losing key architecture characteristics.