[Release] SOTA GGUFs for Qwen3.8-27B: GSQ-RCO at 2.5 to 3.0 bpw
热度趋势
百分比基于当前可用热度信号，而非评论数或独立用户人数。
Qwen3.8-27B的SOTA GGUF模型已发布，采用了最新的GSQ + RCO量化方法。这些模型在保持相同文件大小的同时，提供了更高的质量，并且搜索和量化器都经过了学习。其中，3.00 bpw版本（10.1 GB）在AIME25上与基础模型表现一致（100.00），在GPQA-Diamond（88.89 vs 89.90）和LiveCodeBench v6（84.57 vs 85.71）上的表现也与基础模型非常接近。开发团队欢迎用户提供反馈，并提出希望量化的模型请求。
We're releasing Qwen3.8-27B quantized with our newest methods, GSQ + RCO. Higher-quality models, same file size, now with the search and the quantizer both learned.
- GSQ (Gumbel-Softmax Quantization): post-training scalar quantization that jointly learns grid assignments and scales, closing most of the scalar-vs-vector gap at 2 to 3 bits while staying fully GGUF-deployable
- RCO (Riemannian Constrained Optimization): assigns a quantization type to every tensor under a strict size budget by gradient descent directly on the task loss, no per-constraint tuning
- Three GGUFs (2.50 / 2.75 / 3.00 bpw, 8.4 to 10.1 GB) plus the vision projector
Uniform quants give every tensor the same precision. RCO finds where the precision actually matters and GSQ quantizes every tensor at its assigned type. The result runs unmodified in llama.cpp, Ollama, and LM Studio.
Results (vs. the BF16 base and Unsloth Dynamic quants) To the best of our knowledge, these are the best size-to-accuracy GGUFs available for Qwen3.8-27B at these file sizes: at every size we evaluated, they match or beat the strongest published quantizations.
- 3.00 bpw (10.1 GB): matches the base model on AIME25 (100.00), within about one point on GPQA-Diamond (88.89 vs 89.90) and LiveCodeBench v6 (84.57 vs 85.71)
- 2.75 bpw (9.3 GB): AIME25 100.00, and its zero-shot average actually exceeds BF16 (75.70 vs 74.34)
- At matched ~8.4 GB size: +10.0 AIME25, +8.6 GPQA-Diamond, +4.6 LiveCodeBench over UD-IQ2_S
- GSQ: paper https://arxiv.org/abs/2604.18556 | code https://github.com/IST-DASLab/GSQ
- RCO: paper https://arxiv.org/abs/2605.00649 | code https://github.com/IST-DASLab/RCO
This is the first of a series: we plan to release more SOTA GGUFs across model families as a lab (ISTA Deep Algorithms and Systems Lab).