Release of Compressed DeepSeek-V4.1-Flash-1017 Models
The ISTA Deep Algorithms and Systems Lab has announced the release of a specialized “Coder” build of the DeepSeek-V4.1-Flash-1017 model. This version utilizes Gumbel-Softmax Quantization (GSQ) and Riemannian Constrained Optimization (RCO) to achieve significant compression. The released Coder build is a sparse mixture-of-experts (MoE) model containing 763.2 billion parameters, of which 551.9 billion are in the transformer.
According to documentation provided by ISTA, the Coder build is an expert-pruned GGUF version that retains 192 of the original 384 routed experts per layer. This pruning reduces the transformer weight size to 72.1 GB, compared to 138.2 GB in the unpruned IQ2_XXS build. The compression achieves a ratio of approximately 1 bit per parameter relative to the original model.
The lab noted that RCO was employed to select which experts to retain by optimizing the KL divergence to the unpruned model. This method enforces a strict budget of experts per layer to maintain compatibility with the GGUF storage format.
Performance Benchmarks and Discrepancies
Initial performance data released by ISTA suggests the pruned Coder model retains a high percentage of the base model’s capabilities. In LiveCodeBench v6, the model scored 88.57 against the base model’s 89.71, representing a 98.7% retention rate. On the SWE-bench Verified benchmark, which tests agentic trajectories in real repositories, the model scored 77.00 compared to the base model’s 81.20, a 94.8% retention.
However, subsequent benchmarking conducted by Unsloth has challenged some of these findings. In a report released by Unsloth, the company claimed that their DeepSeek UD version 3.1 GGUFs outperformed the GSQ-RCO versions by 10–20% when tested at extended context lengths (512K for agentic tasks and 128K for general tasks).
Unsloth’s data, which utilized significantly larger datasets—including 35.1 billion tokens compared to the 1.3 billion used in previous runs—suggested that ISTA’s claimed recovery rates were overstated. Specifically, Unsloth reported that the GSQ-RCO IQ2_XXS model, which ISTA claimed had a 99.80% recovery rate, actually measured at 98.38% in their tests. Unsloth characterized the higher claim as “nearly impossible” based on their measurements.
Uncensored Variations and Hardware Requirements
The AtomicChat team has also released an “uncensored” version of the DeepSeek-V4.1-Flash-1017 GSQ-RCO quantizations. According to the team, this version was modified by adjusting 82 dense tensors to reduce refusals. Their testing on a 12 GB VRAM setup showed that refusals dropped from approximately 85% to 0% across various model sizes, including the IQ1_M, IQ2_XXS, and IQ2_XS versions.
The team reported that the IQ2_XXS version maintained high performance in reasoning tasks, with an MMLU-500 score of 82.8% and a negligible KL divergence compared to the stock model.
The deployment of these large-scale models remains hardware-intensive. While the compressed versions are designed for lower VRAM environments, users have noted that running these models typically requires significant system RAM and high-speed SSD storage to manage the large Engram files.