Qwen3.8-27B SOTA GGUF Released

IST-DASLab released Qwen3.8-27B GGUF models applying new quantization technologies: GSQ (Gumbel-Softmax Quantization) and RCO (Riemannian Constrained Optimization). These models can run without modification on major inference engines such as llama.cpp, Ollama, and LM Studio.

Core Technologies and Performance

• GSQ: A learnable scalar quantization technique that reduces the performance gap between scalar and vector quantization in the 2-3 bit range.
• RCO: Assigns quantization types based on per-tensor importance to minimize task loss within capacity constraints.

Benchmark Results

• 3.00 bpw (10.1 GB): Recorded performance identical to the original BF16 model on AIME25 (100.00), and showed approximations within 1 point of the original model on GPQA-Diamond and LiveCodeBench v6.
• 2.75 bpw (9.3 GB): Achieved an AIME25 score of 100.00, with a zero-shot average score (75.70) exceeding that of the BF16 original model (74.34).
• Comparative Advantage: Showed performance improvements of +10.0 points on AIME25, +8.6 points on GPQA-Diamond, and +4.6 points on LiveCodeBench compared to Unsloth Dynamic (IQ2_S) at approximately 8.4 GB.