ISTA-DASLab Releases GSQ-RCO Quantized GGUFs for Qwen3.8-Flash-Next and Expert-Pruned Coder Build

ISTA-DASLab has released GSQ-RCO quantized GGUFs for Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 176.9B parameters. The release includes four standard quantized variants ranging from 2.40 to 3.50 bpw and a specialized Coder build that removes half of the model's routed experts.

Quantization Techniques

The release utilizes two primary optimization methods:

• GSQ (Gumbel-Softmax Quantization): A post-training scalar quantization method that learns grid assignments and group scales to close the gap between scalar and vector quantization at low bit-widths.
• RCO (Riemannian Constrained Optimization): Enforces exact memory budgets via gradient descent on task loss, used here to assign quantization types and select experts for retention.

Performance Metrics

The IQ3_S variant (3.50 bpw, 83.6 GB) matches or exceeds the BF16 base model on key benchmarks:

• AIME25: 100.00
• GPQA-Diamond: 92.93 (vs 91.92 for BF16)
• LiveCodeBench v6: 86.86 (vs 87.43 for BF16)
• Task Average: 93.26 (vs 93.12 for BF16)

The Q2_0 variant (2.40 bpw, 66.4 GB) achieves a zero-shot average of 78.00, surpassing the BF16 value of 76.94 at approximately one-fifth the size.

Expert-Pruned Coder Build

The Coder build removes 256 of 512 routed experts per layer, selected by RCO optimizing KL divergence. Retained weights remain at 3.5 bpw.

• Effective Bitwidth: The combined pruning and quantization result in an average of 1.89 bits per parameter of the original transformer.
• Memory Footprint: The total file size is 58.4 GB, but the resident working set is only 29.6 GB, allowing it to run on a single 32 GB accelerator.
• Capability Retention: SWE-bench Verified score is 75.60 (91.3% of BF16), and LiveCodeBench v6 is 86.28 (98.7% of BF16).