Unsloth released GGUF versions of DeepSeek-V4.1-Flash-1017 Dynamic v3.1

On October 30, 2026, Unsloth posted GGUF versions (quantized files for local execution) of the DeepSeek-V4.1-1017 Dynamic V3.1 model in preview status. In a Reddit post, the team claims that its dynamic quants retain greater accuracy at the same size than ISTA's counterparts (GSQ-RCO). Unsloth compared both lineups on seven benchmarks, including AIME25, LiveCodeBench, GPQA Diamond, and Aider Polyglot.

The difference from previous releases lies in the comparison methodology and the reproducibility claim. Unsloth writes that it could not reproduce ISTA's results even when using their best figures. The most noticeable discrepancy is on AIME25: ISTA reported 100% accuracy recovery relative to the base model, while Unsloth's tests yielded about 91.88%. This does not mean ISTA's quants are bad — rather, that the stated numbers were not confirmed in an independent run. Unsloth also notes that its previous quant versions are inferior to the new Dynamic V3.1 on almost all metrics.

In practice, this means that when choosing a GGUF for local inference, one should look not only at file size but also at how quantization affects specific tasks. Dynamic quantization (different precision for different layers) can yield gains on some benchmarks and losses on others. Users for whom reproducibility matters should run their own tests rather than rely only on the tables from the release. It is also worth considering that the model is marked as preview — behavior may change in the final version.

What remains unknown: the full test configurations and the hardware on which the measurements were taken are not disclosed in the post. In addition, it is unclear when the stable version of DeepSeek-V4.1-1017 Dynamic V3.1 will be released and whether the results will be revised after feedback from ISTA.