VibeusUnsloth has released preview Dynamic V3.1 GGUFs for... - Vibeus
Unsloth has released preview Dynamic V3.1 GGUFs for DeepSeek-V4.1-Flash-1017, alongside a size-versus-quality comparison across seven benchmarks.
The team could not reproduce ISTA’s reported GSQ-RCO results, even using its highest observed scores. The largest discrepancy was AIME25: a claimed 100% recovery to the baseline versus roughly 91.88% in Unsloth’s tests.
In Unsloth’s measurements, Dynamic V3.1 leads GSQ-RCO on nearly all metrics across AIME25, LiveCodeBench, GPQA Diamond, Aider Polyglot and the other reported tests. The older Dynamic v3 also leads on nearly all of these comparisons. These are the releasing team’s benchmark results.
Each benchmark was repeated at least three times and the scores averaged; AIME25 was run sixteen times. The methodology and additional plots are in the linked source’s accompanying blog, with more results planned for next week.
The comparison also covers KL divergence and top-1 agreement across coding, agentic coding, chat and other domains. Quantization aims to reduce both the divergence from the original model and the storage required. At KLD=0 the logits would match the 16-bit model exactly. Unsloth reports better KLD and top-1 results for v3.1 than GSQ-RCO in every domain tested.
Lower-bit versions still reason somewhat longer than the unquantized model. The team plans to investigate this further; its current comparisons also include reasoning length and plots for truncated answers.
For local use, the post mentions 128 GB RAM or unified-memory devices, and a 64 GB setup using engram offloading. The attached chart separately lists RAM + VRAM or unified memory with Engram on SSD: 128–145 GB for 1-bit, 176–205 GB for 2-bit, 256–290 GB for 3-bit, 305 GB for 4-bit and 315 GB for 8-bit.
The source’s numerical table, in order: size GB | AIME25 | HMMT Feb26 | GPQA-D | MMLU-Pro | IFEval | LCB v6 at 32K | average of six | Aider pass@1. A dash means no value is given.
UD-Q8_K_XL: 508.4 | 97.08 | 68.94 | 90.47 | 88.61 | 95.10 | 85.90 | 87.68 | —
UD-Q2_K_XL: 334.8 | 97.50 | 69.89 | 89.58 | 88.28 | 94.92 | 84.76 | 87.49 | —
UD-IQ2_XXS: 300.5 | 96.88 | 67.61 | 88.95 | 87.53 | 95.19 | 83.62 | 86.63 | 44.38
UD-IQ1_M: 250.4 | 95.00 | 62.88 | 88.38 | 87.53 | 95.10 | 84.38 | 85.55 | 42.13
UD-IQ1_S: 233.0 | 92.92 | 57.01 | 87.63 | 86.62 | 94.92 | 82.10 | 83.53 | 36.52
GSQ-RCO IQ2_XS: 265.8 | 95.00 | 64.39 | 88.64 | 86.62 | 94.27 | 84.95 | 85.64 | 41.57
GSQ-RCO IQ2_XXS: 248.8 | 95.21 | 57.95 | 85.80 | 86.53 | 93.72 | 82.86 | 83.68 | 36.52
GSQ-RCO IQ1_M: 231.3 | 91.88 | 53.60 | 85.10 | 85.49 | 91.77 | 82.86 | 81.78 | 29.78
Unsloth Desktop is also getting Browser Use, automatic context compaction, a new text-to-speech pipeline with voice cloning and separation, sandboxing and background customization. The original announcement includes links to the model files, documentation, benchmark methods and Desktop repository.
