ISTA releases pruned, quantized GGUF coder build of DeepSeek V4.1 Flash 1017

ISTA's DAS Lab released an experimental coding build of DeepSeek-V4.1-Flash-1017. It removes half of the model's routed experts and stores the rest at very low precision. The transformer weights drop to 72.1 GB from 138.2 GB for the unpruned IQ2_XXS build. The base model has 763.2B parameters, 551.9B of them in the transformer, with 384 routed experts per layer and 6 active per token. The lab's model card reports 88.57 on LiveCodeBench v6 against a base of 89.71, which it puts at 98.7% retained. It also reports 77.00 on SWE-bench Verified against 81.20, or 94.8% retained.

The two techniques do different jobs. Expert pruning deletes whole feed-forward blocks, so it shrinks both the file and resident memory and narrows the router's candidate pool. Quantization keeps every remaining parameter but stores it at lower precision. The build combines them. 192 of 384 experts are removed in each layer, and the survivors keep the per-tensor quantization types of the IQ2_XXS release, which the card gives as 2.00 bits per weight. According to the card, the kept 280.1B parameters average 2.06 bits, because the dense tensors held at higher precision make up a larger share of what remains.

The card divides the 72.12 GB of transformer weights by the original 551.9B transformer parameters to get 1.05 bits, so every removed expert counts as zero bits. Measured over the whole model, including the Engram lookup tables, the card gives 1.92 bits per parameter. The Engram tables are 110.6 GB, are shared with the unpruned release, and are read one row per token. The card says they may stay on disk.

The choice of which experts to delete is handled by RCO. The card describes the method as optimizing the KL divergence between the pruned and unpruned model on calibration data, rather than ranking experts by a heuristic score. The search runs on the IQ2_XXS build itself, so it removes the experts whose absence changes that build's outputs least. The card notes a constraint of the GGUF format: it stores one expert count for the whole model, so each layer must keep the same number of experts. RCO enforces a separate budget for each layer to meet that requirement. Because experts specialise, the calibration data determines which capabilities survive. The card says this release targets code, agentic tool use, long context and vision. It also says degradation outside those areas is an accepted cost, and it recommends the unpruned GSQ-RCO releases for general-purpose use.

The benchmark numbers come from the lab's own measurements at reasoning effort 50. The card offers only two coding benchmarks and one informal check. The check is a generated SVG of a pelican riding a bicycle, which the card describes as animated. The card explains the gap between the two benchmarks by arguing that multi-step agentic tasks accumulate errors across turns, so they tolerate capacity loss less well than single-problem code generation. The card offers this as an expectation, not a measurement. It doesn't say how many runs were averaged or what variance the scores carry.

Discussion on Reddit's LocalLLaMA community took a more skeptical line, and those points are the commenters' own. One argued that the single-file HTML demos are not meaningful comparisons against the unpruned IQ2_XXS model. Another cited Qwen3.8-27B at 90.3 on LiveCodeBench and Claude Opus 4.6 at 81.42% on SWE-bench Verified. That commenter asked why a much larger pruned model is worth running when smaller models score higher. A third commenter referenced reports that Unsloth or another group had failed to match ISTA's numbers. Others asked for an intermediate variant of 90 to 110 GB, with higher precision or fewer pruned experts, that would suit 128 GB systems better than the 182.72 GB total on disk.

Practical use requires llama.cpp b11870 or newer, which supports DeepSeek-V4.1, or the lab's Strata engine. The card's download instructions assume an Engram file that is symlinked into the shard set. The full set of files totals 182.72 GB, against 248.84 GB for the unpruned build.