ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
AI model · https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
First seen by Oriel on .
AI summary
ISTA-DASLab provides three non-uniform GGUF quantizations of DeepSeek-V4.1-Flash-1017, with IQ2_XXS matching the base model’s task average at 2.00 bits per weight.
- GSQ learns low-bit scalar quantization for each tensor, while RCO assigns tensor-specific quantization types under an exact total-size budget.
- The IQ1_M, IQ2_XXS, and IQ2_XS builds range from 231.32 GB to 265.83 GB, including a shared 110.60 GB Engram file.
- IQ2_XXS scores 93.54 across AIME25, GPQA-Diamond, and LiveCodeBench v6, matching the base model; IQ1_M retains 98.7% of its task average.
- Strata runs every build with a single 12 GB GPU, keeping the backbone in system memory and the Engram tables memory-mapped on disk.
- The GGUFs currently run in Strata; the page says llama.cpp does not yet support DeepSeek-V4.1.
#gguf #model quantization #deepseek #llm inference
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 615, latest observation ; +504 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 2,741,241, latest observation ; +2,708,762 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF | https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF Read discussion |