Unsloth releases DeepSeek-V4.1-Flash-1017 GGUFs and disputes quantization claims

• Dynamic v3.1 preview GGUFs are available for DeepSeek-V4.1-Flash-1017; the release announcement claims better accuracy at comparable sizes across nearly all tested benchmarks.
• Perfect AIME25 scores did not reproduce: the smallest GSQ-RCO build scored 91.88% in Unsloth's comparison, against ISTA's reported 100%.
• Lower-bit quants generated longer reasoning, according to Unsloth's measurements, which also tracked answers cut off at the token limit.
• Follow-up coding tests increased the budget to 128K, with LiveCodeBench and Aider results published in a subsequent update.

The GSQ-RCO model card excludes its 196.6B-parameter Engram tables from the advertised bits-per-weight figures. The smallest Unsloth download still weighs 233 GB, including 110.6 GB of Engram lookup tables that can stay on SSD.

Unsloth advertises 64 GB+ RAM support in its launch post. Its hardware guide describes 64 GB configurations as slower, while listing 128–145 GB of combined RAM/VRAM or unified memory for 1-bit inference with Engram on SSD.

Engram rows load on demand into the OS cache; offloading preserves the full disk requirement. The guide's local execution options are:

• Unsloth Desktop, which configures offloading automatically.
• Unsloth's documented llama.cpp branch, with --load-mode mmap and --lazy-mode on in the SSD-offloading example.

ISTA reports 100% AIME25 accuracy for all three GSQ-RCO builds in its model card. Its IQ2_XXS build also claims 100% recovery of the base model's average across AIME25, GPQA-Diamond and LiveCodeBench v6.

Unsloth challenged those recovery claims in a later reply. Its published comparison separates raw accuracy from relative recovery:

• Unsloth UD-Q8_K_XL reference: 97.08% on AIME25.
• GSQ-RCO IQ1_M: 91.88%, a 5.20-point gap against that reference.
• Relative recovery: 94.64% of Unsloth's reference score.

The baselines differ: ISTA evaluates against the released FP8/FP4 checkpoint; Unsloth's table uses its 508.4 GB UD-Q8_K_XL reference.

Coding also separated similarly sized files. On Aider pass@1, Unsloth's 233 GB IQ1_S scored 36.52%, versus 29.78% for GSQ-RCO's 231.3 GB IQ1_M, a 6.74-point gap in the same published table.

Unsloth says every quant ran through the same evaluation process, with settings, graders and repetitions disclosed in its methodology reply. The methods table specifies:

• Generation settings: temperature 1.0, reasoning effort 50, initially a 32,768-token cap.
• AIME25: 30 problems, 16 samples each, totaling 480 answers per quant.
• GPQA-Diamond: eight samples per question.
• LiveCodeBench v6: three samples per problem, with sandboxed execution grading.
• Aggregate: six benchmarks; Aider is reported separately.

ISTA attributed the disagreement to its 128K output budget versus Unsloth's initial 32K budget, according to a summary of its Reddit response. Both teams report reasoning effort 50.

Unsloth also defended KL divergence against claims that it is a poor or gameable quantization metric in its release thread. KL measures drift between next-token probability distributions on the evaluated text; top-1 agreement counts matching most-likely tokens.

After criticism about testing on a single dataset, Unsloth says it evaluated KL and top-1 agreement across multiple held-out sets. Its quantization analysis includes wiki1100 and long17k, alongside the downstream task benchmarks.

Unsloth counted token-cap failures in its results table, the most useful addition to this quant fight. Across the six capped benchmarks, the share of truncated answers was:

• UD-Q8_K_XL reference: 1.48%.
• Unsloth IQ1_S, 233 GB: 4.15%.
• GSQ-RCO IQ1_M, 231.3 GB: 4.69%.

Unsloth reports that lower-bit quants generally reasoned longer, so these capped accuracy results also reflect whether a model finished its answer. Smaller quants can nevertheless yield more tokens per second when they fit the hardware better, Unsloth noted in a reply about throughput.

Unsloth said its initial 128K checks preserved its conclusion in an early reply, then published the coding-specific follow-up below. LiveCodeBench's Q8 reference increased from 85.90% at 32K to 88.95% at 128K, +3.05 points.

The Aider follow-up used standard two-attempt pass@2, as Unsloth specified in a separate reply; the earlier table reported pass@1.

AIME25 and six other evaluations were still running when Unsloth posted its progress update.