AI Briefing
KOSign in

ISTA DASLab Releases 1-bit Expert-Pruned DeepSeek-V4.1-Flash-1017 Coder

·2026.11.02 23:33

Key point

The expert-pruned build retains 98.7% of LiveCodeBench v6 performance while reducing transformer weights to 72.1 GB.

Details

  • ISTA DASLab released a Coder build of DeepSeek-V4.1-Flash-1017 using GSQ and RCO, compressing the 763.2B parameter model to 1.05 bits per parameter.
  • The build prunes half of the routed experts (192 of 384 per layer), reducing transformer weights from 138.2 GB to 72.1 GB.
  • LiveCodeBench v6 score is 88.57, retaining 98.7% of the base model's 89.71 score.
  • SWE-bench Verified score is 77.00, retaining 94.8% of the base model's 81.20 score.
  • GSQ (Gumbel-Softmax Quantization) jointly learns grid assignments and group scales to close the gap between scalar and vector quantization.
  • RCO (Riemannian Constrained Optimization) enforces exact expert budgets per layer via gradient descent on KL divergence to the unpruned model.
  • The model requires llama.cpp b11870 or newer and runs in Strata; shared Engram and vision projector files are identical to the unpruned release.
quantizationdeepseekggufexpert-pruningllm-compressionista-daslab
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditISTA Releases GSQ-RCO Quantized GGUFs for DeepSeek-V4.1-Flash-1017RedditUnsloth claims 10-20% performance lead over GSQ-RCO on DeepSeek-V4.1 GGUF benchmarksHacker NewsSamsung Labs Releases LittleBit-2 for Sub-1-Bit LLM CompressionPyTorchKRHugging Face Adds Native GGUF Execution to TransformersTLDR AIDeepSeek Open-Sources DeepInfer Framework for V4.1 Flash InferenceDeepSeekDeepSeek Releases Search and Fetch APIs... $0.75 per 1,000 Requests