The ISTA Deep Algorithms and Systems Lab has released a Coder build of DeepSeek-V4.1-Flash-1017 that removes half of the routed experts and brings transformer weights to 1 bit per parameter, counted against the 551.9 billion parameters of the original model. The release matters for anyone running models locally: the GGUF works with llama.cpp from build b11870 and with Strata, and the repository ships the complete per-tensor RCO allocation.

The starting point is a sparse mixture-of-experts model: 384 routed experts per layer, 6 active per token, 763.2 billion total parameters, of which 551.9 billion are in the transformer. The pruned build keeps 192 of the 384 experts in every layer and cuts transformer weights from 138.2 GB in the unpruned IQ2_XXS build to 72.1 GB. The 1-bit-per-parameter figure comes from dividing 72.12 GB by the original 551.9 billion transformer parameters; removed experts count as zero bits, while the 280.1 billion kept parameters average 2.06 bits. No individual weight is stored below 1 bit.

For anyone considering a self-hosted deployment, however, that number does not tell the whole story. The shared Engram file of 110.6 GB and the vision projector remain identical to the unpruned release. The transformer weight reduction frees 66.1 GB, but the Engram file alone weighs 110.6 GB, more than the 72.1 GB of pruned transformer weights. If you already have the unpruned build, the shared files do not need to be downloaded again; if you start from scratch, the total footprint does not shrink in proportion to the advertised average bitrate.

Two distinct techniques sit behind the result. GSQ is a post-training scalar quantization method that jointly learns grid assignments and group scales, narrowing most of the gap between scalar and vector quantization at low bit-widths while remaining deployable in standard GGUF types. RCO imposes exact budgets through gradient descent on the task loss, without per-constraint tuning: here it selects which experts to keep by optimizing the KL divergence to the unpruned model. The uniform per-layer constraint is not a minor detail. GGUF stores a single expert count for the whole model, so every layer must keep exactly the same number of experts; uneven layers would require padding that gives back the memory saved by pruning. RCO treats each layer's budget as a hard constraint and holds it exactly at every optimization step, whereas a penalty formulation would only meet the budgets approximately and would need a tuned coefficient per layer.

On benchmarks the drop is small for single-shot code generation: LiveCodeBench v6 falls from 89.71 to 88.57, retaining 98.7% of capacity. On SWE-bench Verified the decline is sharper: 77.00 versus 81.20, or 94.8%. The explanation lies in the nature of the test: SWE-bench Verified runs long agentic trajectories over real repositories and tolerates capacity reduction less well because errors accumulate across turns. The lab also published real examples, each in a single HTML file: an animated SVG of a pelican riding a bicycle, a three.js voxel pagoda garden, and a playable first-person voxel sandbox. Those creative capabilities survive pruning, but the release remains experimental, and the team asks for feedback especially on capabilities not represented in the calibration mixture.

The publication of the per-tensor RCO allocation provides a transparent basis for understanding which experts were sacrificed. For users already running the unpruned build, the immediate benefit is that only the new transformer weights need to be downloaded, while the shared files stay in place. For those starting from zero, the real saving must be measured across the whole model directory, not just against the average bitrate of the retained parameters.