ISTA发布DeepSeek-V4.1-Flash编码模型：GSQ+RCO压缩

We have released a Coder build of DeepSeek-V4.1-Flash-1017, compressed with GSQ and RCO, in which half of the model's routed experts have been removed.

DeepSeek-V4.1-Flash-1017 is a sparse mixture-of-experts model: 384 routed experts per layer, of which 6 are active per token, 763.2B parameters, 551.9B of them in the transformer.

What's inside

• One expert-pruned GGUF: 192 of 384 experts kept in every layer, 72.1 GB of transformer weights, against 138.2 GB for our unpruned IQ2_XXS build.
• 1 bit per parameter, counted against the original model: 72.12 GB divided by the original 551.9B transformer parameters gives 1.05 bits, which we round to 1. The removed experts count as zero bits. No individual weight is stored below 1 bit: the 280.1B parameters the Coder keeps average 2.06 bits.
• The shared Engram file (110.6 GB) and vision projector are the same files as in the unpruned release, so they need to be downloaded only once.
• GSQ (Gumbel-Softmax Quantization): post-training scalar quantization that jointly learns grid assignments and group scales, closing most of the gap between scalar and vector quantization at low bit-widths while remaining deployable in standard GGUF types.
• RCO (Riemannian Constrained Optimization): enforces exact budgets by gradient descent on the task loss, without per-constraint tuning. Here it selects which experts to keep, optimising the KL divergence to the unpruned model, under one exact budget per layer.

Why one budget per layer

GGUF stores a single expert count for the whole model, so every layer must keep exactly the same number of experts; uneven layers would need padding that gives back the memory pruning saved. RCO treats each layer's budget as a hard constraint and holds all of them exactly at every optimisation step, while the layers are still optimised jointly through one loss on the full model. A penalty formulation would only meet the budgets approximately and would need one tuned coefficient per layer.

Results

• LiveCodeBench v6: 88.57 against 89.71 for the base model, retaining 98.7%
• SWE-bench Verified: 77.00 against 81.20, retaining 94.8%

SWE-bench Verified runs long agentic trajectories over real repositories, and tolerates the capacity reduction less well than single-problem code generation, as expected when errors accumulate across turns.

Real-world examples, each a single HTML file written by the Coder: an animated SVG of a pelican riding a bicycle, a three.js voxel pagoda garden, and a playable first-person voxel sandbox. For full post incl. videos see: https://x.com/mhelcig/status/2117257900863156518

Links

• Coder (expert-pruned): https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF
• Quantized models (unpruned): https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
• GSQ: paper https://arxiv.org/abs/2604.18556 | code https://github.com/IST-DASLab/GSQ
• RCO: paper https://arxiv.org/abs/2605.00649 | code https://github.com/IST-DASLab/RCO

The repository ships the complete per-tensor RCO allocation. The GGUF runs in llama.cpp b11870 or newer and in Strata.

The Coder build is an experimental release and feedback is welcome, particularly on capabilities that were not represented in the calibration mixture. Requests for models to quantize or prune are also welcome.

From the ISTA Deep Algorithms and Systems Lab.