ISTA-DASLab Ships 2-Bit DeepSeek V4.1 Flash at 100% Task Recovery

A 2-bit scalar quantization of DeepSeek V4.1 Flash matches its base model exactly across three reasoning benchmarks. ISTA-DASLab's IQ2_XXS build on Hugging Face, at 2.00 bits per weight and 248.84 GB on disk, posts 100.00 on AIME25, 90.91 on GPQA-Diamond, and 89.71 on LiveCodeBench v6, the same three scores as the FP8-dense/FP4-routed reference measured at reasoning effort 50.

The underlying model is a 763.2B-parameter sparse MoE with 384 routed experts per layer, 6 active per token, and a 1M-token context window. The smaller IQ1_M build (1.75 bpw, 231.32 GB) holds 98.7% of the base task average; the 2.25 bpw IQ2_XS lands at 99.8%.

The method pairs two components. GSQ is described in the card as post-training scalar quantization that jointly learns per-coordinate grid assignments and per-group scales via a Gumbel-Softmax relaxation, which, the lab writes, "GSQ closes most of the gap between scalar and vector quantization at 2 to 3 bits" while staying inside the standard GGUF format. RCO then "Assigns one of K quantization types to each of N tensors under a total size budget," using a Riemannian manifold reformulation that enforces the budget exactly without extra hyperparameter tuning.

The hardware pitch is where it gets strange. "With Strata, the backbone does not need to fit in VRAM either: every build runs on a 12 GB GPU, with the rest of the model held in system memory," the card states. The 110.60 GB Engram lookup tables, which hold 196.6B of the model's 763.2B parameters, stay memory-mapped on an SSD and are read one row per token. llama.cpp support requires release b11870 or newer, dated October 25, 2026. The page shows 1,105,751 downloads in the last month.

It lands amid a dense run of open-weight quantization work: our tracker logged 353 open-source AI stories in the last 90 days.