DASLab has published compressed DeepSeek-V4.1-Flash-1017 weights that it says run through Strata on a single 12 GB GPU. The smallest set still totals 231.32 GB. For developers considering local inference, GPU memory is only part of the hardware bill.
Strata 0.2.5, released October 25, 2026, brought DeepSeek into its desktop app's model picker. That makes the weights easier to select, but the release also introduced a speed setting that changes how many experts the model uses. This article examines the published documentation and results; we haven't downloaded the weights or reproduced inference.
Budget for the whole download
The model card lists three builds. The recommended IQ2_XXS set totals 248.84 GB, including a shared 110.60 GB Engram lookup file. Those lookup tables can stay on an SSD rather than occupy RAM or VRAM.
Strata's DeepSeek guide explains the larger installation figures: it also downloads an approximately 8 GB DSpark draft model. Its rounded totals are 239 GB for the smallest set and 257 GB for the middle option. Those figures describe a different bundle from the weights alone, so the two download tables aren't interchangeable.
RAM affects speed too. The guide says configurations below about 40 GB are untested and slow. You shouldn't read the 12 GB GPU claim as a complete minimum specification.
Two-bit weights don't mean identical execution
The compression builds on GSQ research, first submitted in April. GSQ optimizes how individual weights map onto a small set of numerical values. The researchers describe retaining compatibility with existing scalar inference kernels and applying the method to GGUF checkpoints. The underlying method therefore predates this DeepSeek release.
DASLab reports that IQ2_XXS matches the base model on AIME25, GPQA-Diamond and LiveCodeBench v6 at reasoning effort 50. That is a three-benchmark result, not a measurement of your coding agent's success rate.
Runtime choices add another variable. The 0.2.5 release notes say Faster mode uses five of the model's six experts per token and defaults to that mode on PCs with 16 GB of graphics memory or less. Full quality uses all six. A speed comparison needs to record that setting before its results can be compared with another run.
Check the configuration behind the speed
Strata's README reports 4 to 6 output tokens per second for the 2.00-bit set on an RTX 5070 with 12 GB VRAM and 64 GB RAM, measured using Strata 0.2.5. These are the project's measurements. Its prompt-processing rate is listed separately, which matters if an agent repeatedly submits long inputs before receiving an answer.
The next useful evidence is an independent run that records the quantization, expert setting and reasoning effort alongside task correctness and speed. Until then, the documented storage and runtime requirements give prospective users a concrete first check: whether their whole computer can support the setup before they begin a download exceeding 200 GB.