ISTA releases 763B DeepSeek coder at 1 bit per param, pruning half the experts with 95% performance kept
Loginhe · reddit · 2026-11-02
ISTA DASLab released a GSQ-RCO Coder build of DeepSeek-V4.1-Flash-1017, a sparse MoE with 763.2B params (384 routed experts per layer, 6 active). Highlights:
- Expert pruning: keeps 192/384 experts per layer; transformer weights shrink to 72.1 GB vs 138.2 GB unpruned (IQ2XXS), 1.05 bits per original parameter; the retained 280.1B params average 2.06 bits.
- Techniques: GSQ (Gumbel-Softmax Quantization) jointly learns grid assignments and group scales; RCO (Riemannian Constrained Optimization) enforces exact per-layer budgets via gradient descent on task loss, optimizing KL divergence to the unpruned model (GGUF requires uniform expert counts).
- Results: LiveCodeBench v6 88.57 vs 89.71 base (98.7% kept); SWE-bench Verified 77.00 vs 81.20 (94.8%), with longer agentic trajectories degrading more.
- Shared Engram file (110.6 GB) reused; runs on llama.cpp b11870+ and Strata. Papers and code for GSQ/RCO are open source.
More from Infra
- Musk: Superintelligence compute will "round up to 100%" be in space — inductionheads · 2026-11-02
- Liquid AI launches Pipette, an on-device LLM benchmark with live leaderboard — JosephJacks_ · 2026-11-02
- T3 Code claims 2x OpenAI's token throughput — mocked for conflating routing with inference — scaling01 · 2026-11-02
- CI tip: GitHub Actions rounds each job up to a full minute — merge short jobs to save cost — DanielLockyer · 2026-11-02
- UBS: Musk's Terafab rethinks fabs — wafers spend only ~25% of cycle time processing — BenBajarin · 2026-11-02
- Unsloth's 1-bit DeepSeek-V4.1-Flash GGUF beats other quants by 10%+ on agentic benchmarks — solyarisoftware · 2026-11-02