DASLab releases GSQ-RCO GGUFs of DeepSeek-V4.1-Flash-1017
Original title: ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF — new model trending #29 on Hugging Face
The IQ2_XXS build matches the base model’s reported task average, and every build is said to run on a 12 GB GPU with Strata.
ISTA’s DASLab published non-uniform GGUF quantizations of DeepSeek-V4.1-Flash-1017, a 763.2B-parameter mixture-of-experts model with 384 routed experts and a 1M-token context. GSQ handles low-bit scalar quantization and RCO assigns per-tensor types under a size budget. Builds are IQ1_M (1.75 bpw), IQ2_XXS (2.00), and IQ2_XS (2.25), with a shared Engram file at IQ4_NL. The card says IQ2_XXS reaches 100.0% of the base model’s task average of 93.54.
Unconfirmed: The model card marks GPU memory placement for the 12 GB Strata setup as still to be confirmed.
- Backbone sizes are 120.72 GB, 138.24 GB, and 155.23 GB.
- The shared Engram file is 110.60 GB and is identical for every build.
- IQ1_M trades 1.2 task-average points for 17.5 GB less storage.
- Six of 384 routed experts are active per token.
- Organizations
- DASLabInstitute of Science and Technology Austria
- AI models
- DeepSeek-V4.1-Flash-1017
From the original article
7 min read at huggingface.cotext-generation · 32,479 downloads · 153 likes · license mit
By ISTA-DASLab · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.