Unsloth releases Qwen3.8-27B Dynamic V3.1 GGUFs with 5-10% accuracy gains

·2026.11.12 01:17

Key point

New 2-bit to 6-bit quantizations achieve higher accuracy on agentic and coding benchmarks using 256K and 128K context windows.

Unsloth released Dynamic V3.1 GGUFs for Qwen3.8-27B, reporting 5-10% accuracy improvements over UD3, GSQ-RCO, and Exllama-3 on DeepSWE and Terminal Bench.
Updated quantizations cover 2-bit through 6-bit sizes; 3-bit quants were not updated as UD3 was found to perform on par with UD3.1 at that size.
Benchmarks used 256K context for agentic tasks and 128K for others, with specific sampling parameters including temperature 1.0 and xhigh reasoning.
Analysis of competitor data approximates that GSQ-RCO used a 96K context budget, where UD3 outperforms at 128K, and Unsloth estimates ISTA's reported 100% accuracy recovery on AIME25 was based on a single trial compared to Unsloth's 14 trials.

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.