AI

[Release] ISTA GSQ-RCO GGUFs for DeepSeek-V4.1-Flash-1017

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 24/10/2026, 19:47:25 UTC.

Briefings

The ISTA Deep Algorithms and Systems Lab released three GGUF quantized versions of DeepSeek-V4.1-Flash-1017, achieving 100% accuracy on AIME25 and GPQA-Diamond at 2.00 bits per weight. The model features 384 routed experts with 6 active per token, 763.2B parameters, and a 1M-token context window, running on a single 12 GB GPU via Strata. Key metrics show IQ1_M (1.75 bpw) scores 92.32% on task average, while IQ2_XXS (2.00 bpw) matched the base model at 93.54%.

Key points

• Three GGUF quantized versions (IQ1_M, IQ2_XXS, IQ2_XS) of DeepSeek-V4.1-Flash-1017 achieved 100% accuracy on AIME25 and GPQA-Diamond at 2.00 bits per weight.
• The model uses 384 routed experts with 6 active per token, 763.2B parameters, and a 1M-token context window, running on a single 12 GB GPU via Strata.

Context

The model is a sparse mixture-of-experts architecture released with FP8 dense weights and FP4 routed experts. It includes a shared Engram file for two lookup tables at IQ4_NL, identical for every build. Results were measured at reasoning effort 50, with IQ1_M matching the base model on every benchmark evaluated.

Limitations

Running the model in llama.cpp is not supported yet, so GGUFs do not run there for now. Results are limited to reasoning effort 50, and IQ2_XS results are pending. The bit-width is calculated as the average over 551.9B backbone parameters.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 24/10/2026, 19:55:29 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibxhztut4phuv74mlzfhrapotxirq4k5r7zgr7wvr2lsecmh4dq2y