ISTA-DASLab's 2-bit DeepSeek V4.1 Flash matches base model exactly

ISTA-DASLab's release comes amid a dense run of open-weight quantization work, with 353 open-source AI stories logged in the last 90 days. The 2-bit build's exact match on three reasoning benchmarks suggests that aggressive scalar quantization can preserve model accuracy at 2 bits per weight, a notable result for the field.

The method combines GSQ, a post-training scalar quantization that learns per-coordinate grid assignments and per-group scales, with RCO, which assigns quantization types to tensors under a total size budget. The lab states that GSQ closes most of the gap between scalar and vector quantization at 2 to 3 bits while staying within the standard GGUF format, and RCO enforces the budget exactly without extra hyperparameter tuning.

The page shows 1,105,751 downloads in the last month, indicating strong interest. The smaller IQ1_M build at 1.75 bpw holds 98.7% of the base task average, while the 2.25 bpw IQ2_XS lands at 99.8%, offering trade-offs between size and accuracy.