A 763B-parameter model can be compressed to roughly 2 bits per backbone weight without simply applying one uniform quantization level everywhere. The key is to spend precision where it matters: GSQ optimizes low-bit tensor quantization, while RCO searches for a per-tensor allocation that meets a size budget. In the reported tests, the IQ2_XXS build matches the base model across three benchmarks, though the resulting files still total about 249 GB.
AI Model2026
DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF
Provides non-uniform GGUF quantizations of DeepSeek-V4.1-Flash-1017, using GSQ and RCO to allocate precision per tensor under a size budget. The IQ2_XXS build matches the base model on the three reported benchmarks.
Introduction
Information
- Websitehuggingface.co
- OrganizationsDeep Algorithms and Systems Lab (DASLab), Institute of Science and Technology Austria
- Published date2026/10/23
Categories
More Items
AI Model2026
Analyzes DNA in chat, predicting variant effects, identifying sequence features, and continuing sequences. Its hybrid tokenizer represents DNA as 6-mers alongside ordinary text; it is not clinically validated.
AI Model2026
Provides locally runnable GGUF builds of DeepSeek-V4.1-Flash-1017 with mixed-precision options for different memory budgets. Its refusal behavior is substantially reduced, making it suitable only for controlled safety research and red-team evaluation.
AI Model2026
A sparse MoE language model built for Korean conversation and tool-using tasks, with a context window of up to 262,144 tokens. Its hybrid attention design supports long contexts, and preview weights are released under Apache 2.0.