GSQ-RCO
A mixed quantization strategy for large language models that assigns different bit-widths to different tensors, aiming to achieve a balance between compression ratio and model quality.
Timeline (last 90 days)
Unsloth claims that its UD-series quantized models are almost comprehensively ahead of ISTA's GSQ-RCO quantized versions on multiple benchmarks (AIME25, HMMT Feb26, GPQA-D, MMLU-Pro, IFEval, LiveCodeBench v6, Aider pass@1).
UDv3.1 is clearly better than GSQ-RCO on the KLD and top-1% metrics across multiple domains including coding, agentic coding and dialogue.
Unsloth ran reproduction tests on ISTA's GSQ-RCO GGUFs and failed to obtain the results ISTA reported; on AIME25, ISTA claimed 100% accuracy recovery while Unsloth measured only about 91.88%.
The quantization scheme of Underdog Saluki 27B uses ISTA's GSQ-RCO quantization method.
The scheme uses a GSQ-RCO mixed quantization strategy, allocating bit-widths by tensor, with the IQ3S tier averaging approximately 3.5 bits.
RCO-GSQ is a low-bit quantization scheme proposed by ISTA-DASLab, where GSQ introduces sparsity through weight grouping to reduce precision loss, and RCO performs further compression and correction on the quantization residuals.
IQ2_XXS以约2.06位的平均比特率编码权重，IQ3_S约3.44位，Q4_K_M约4.83位
All Facts (7)
Unsloth claims that its UD-series quantized models are almost comprehensively ahead of ISTA's GSQ-RCO quantized versions on multiple benchmarks (AIME25, HMMT Feb26, GPQA-D, MMLU-Pro, IFEval, LiveCodeBench v6, Aider pass@1).
50%UnverifiedUDv3.1 is clearly better than GSQ-RCO on the KLD and top-1% metrics across multiple domains including coding, agentic coding and dialogue.
50%UnverifiedUnsloth ran reproduction tests on ISTA's GSQ-RCO GGUFs and failed to obtain the results ISTA reported; on AIME25, ISTA claimed 100% accuracy recovery while Unsloth measured only about 91.88%.
50%UnverifiedThe quantization scheme of Underdog Saluki 27B uses ISTA's GSQ-RCO quantization method.
50%UnverifiedThe scheme uses a GSQ-RCO mixed quantization strategy, allocating bit-widths by tensor, with the IQ3S tier averaging approximately 3.5 bits.
50%UnverifiedRCO-GSQ is a low-bit quantization scheme proposed by ISTA-DASLab, where GSQ introduces sparsity through weight grouping to reduce precision loss, and RCO performs further compression and correction on the quantization residuals.
50%UnverifiedIQ2_XXS以约2.06位的平均比特率编码权重，IQ3_S约3.44位，Q4_K_M约4.83位
50%