Unsloth claims 10-20% performance lead over GSQ-RCO on DeepSeek-V4.1 GGUF benchmarks

·2026.11.03 00:57

Key point

Unsloth reports its Dynamic 3.1 GGUFs for DeepSeek-V4.1 outperform GSQ-RCO by 10-20% on agentic tasks at 512K context, while statistical analysis suggests GSQ-RCO's claimed 99.80% recovery for IQ2_XXS is nearly impossible.

Details

Unsloth Dynamic 3.1 GGUFs for DeepSeek-V4.1 outperform GSQ-RCO by 10-20% on DeepSWE, Terminal Bench, and Automation Bench at 512K context.
Statistical re-evaluation of GSQ-RCO's quantization claims indicates IQ1_M recovery is significantly lower than the claimed 98.7% (measured 97.24%), and IQ2_XXS 99.80% recovery is statistically nearly impossible (p=0.0001).
The study utilized 35.1B tokens of benchmarking data, a 27x increase over previous runs, to reduce noise and improve reliability.
Benchmarks included AIME2025 (64x runs), DeepSWE, and Terminal Bench, with specific infrastructure fixes and official graders applied to address prior methodological issues.

quantization unsloth gguf benchmarks gsq-rco deepseek-v4.1

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.