ISTA-DASLab’s GSQ-RCO builds make DeepSeek-V4.1-Flash-1017 available through Strata on a 12 GB GPU. The smallest published backbone-plus-Engram set is still 231.32 GB. “Runs on 12 GB” describes GPU residency in an offloaded workflow; it does not describe the download, system RAM requirement or an all-in-memory model.
EyesTech previously covered the 1017 API update and pricing. These quantized files answer a different question: what must a local machine supply when it holds only part of the model in graphics memory?
The shared lookup file changes the storage calculation
The ISTA-DASLab release lists backbone sizes of 120.72 GB for IQ1_M, 138.24 GB for IQ2_XXS and 155.23 GB for IQ2_XS. Each also needs the same 110.60 GB Engram lookup file. Totals are 231.32, 248.84 and 265.83 GB respectively, using decimal gigabytes.
For the middle build, the shared file is about 44.4% of the complete set: 110.60 ÷ 248.84. Moving from IQ2_XXS to IQ1_M saves 17.52 GB, around 7.0% of that complete footprint. This is EyesTech arithmetic from the listed sizes. A smaller backbone does not shrink the fixed lookup file.
The repository keeps Engram memory-mapped on disk and recommends an SSD. Memory mapping provides an addressable file; it does not make storage latency disappear. The practical question is how often a workload needs data that is absent from faster memory, and how quickly the system can supply it.
Strata adds a deployment package around those weights
The current Strata README lists a roughly 257 GB download and about 265 GB of free disk space for its 2.00 bpw option. That exceeds the quantization repository’s 248.84 GB total because the complete setup includes an additional drafter. Use the engine’s package estimate when preparing disk space, rather than treating the quantized model total as the entire installation.
Strata’s own measurements report 4–6 output tokens per second on an RTX 5070 12 GB with 64 GB RAM, and 9–11 on an RTX 3060 12 GB with 46 GB RAM, for the specified Strata 0.2.5 setup. These are project-reported system results. They are not a controlled GPU comparison: the host machines differ. They demonstrate why GPU model and VRAM alone cannot predict the experience.
The DeepSeek engine guide sets an 8,192-token default context and says setups below about 40 GB RAM are untested and slow. It also documents a consequential default: on systems with at most 16 GiB total GPU memory, the faster mode uses five routed experts instead of six. Full quality uses all six. Record this choice before comparing an installed application with a model-card score.
The same guide says this engine version does not provide simultaneous requests or image input. Its wider application’s capabilities should not be assumed to apply to every model backend. Plan the intended request type and concurrency before downloading hundreds of gigabytes.
GSQ and RCO solve different compression problems
GSQ learns scalar grid assignments and group scales through a Gumbel-Softmax relaxation. RCO chooses compression options under an exact size budget while optimizing an objective that need not split into independent per-layer scores. The combination spends precision unevenly rather than applying one uniform bit width everywhere.
That helps explain the release’s size-versus-quality objective. It does not guarantee identical behavior on every task. The quantization card reports that IQ2_XXS matches the base model’s average of 93.54 across AIME25, GPQA-Diamond and LiveCodeBench v6, at reasoning effort 50. This is a three-benchmark result, not a claim that all outputs, languages or long agent sessions remain unchanged.
For local evaluation, separate compression from runtime changes. Compare the same prompts and acceptance checks with fixed reasoning effort, expert count, context limit and tool access. Otherwise a five-expert speed setting can be mistaken for quantization loss, or a changed reasoning budget can be mistaken for a better checkpoint. Preserve output and timing logs rather than relying on an aggregate impression.
GGUF does not settle runtime compatibility
The quantization card’s dated October 23 note says llama.cpp is unsupported. Upstream pull request 31086 was subsequently merged on October 25, adding DeepSeek-V4.1 Flash support. The older blanket statement is therefore insufficient as a description of today’s upstream code.
A merged architecture implementation still does not prove that these exact split GGUF files load correctly, with their template and lookup tables, in every packaged application. This article has not executed that combination. Strata is the explicitly documented route for these builds; verify the selected engine version and exact artifact before treating another GGUF-capable app as interchangeable.
The release makes a large model accessible to smaller GPUs by using the rest of the computer. A sensible readiness check covers free SSD space, host memory, selected engine, context length and expert mode together. The evidence that matters next is a recorded run on that complete configuration with the intended workload.
Reporting note: Prepared with AI assistance from primary sources checked October 27, 2026. Storage percentages are EyesTech calculations. Runtime speeds and benchmark scores are attributed to their publishers; no local model execution was performed. The cover is an original AI-generated conceptual illustration.
