DeepSeek GSQ-RCO GGUF: a 12 GB GPU claim still comes with a large installation

Model Watch · source review October 25, 2026. The ISTA-DASLab repository was created October 23 and updated October 25. AiBrainWorX has not downloaded, run or benchmarked these weights.

New GSQ-RCO GGUF builds of DeepSeek-V4.1-Flash-1017 make a useful distinction clear: a model can be served with a modest GPU without being a small installation. The GPU capacity in a compatibility claim is only one part of the machine you need.

The shared file changes the storage calculation

ISTA-DASLab’s model card separates each quantized backbone from a shared Engram file. Its IQ1_M table lists 120.72 GB for the backbone and 110.60 GB for Engram, totaling 231.32 GB. Those are the card’s decimal file-size figures, not an AiBrainWorX disk measurement.

The Engram component is required alongside the selected backbone. Looking only at a backbone download can therefore understate the installation substantially. Before choosing a build, budget for every required file, any runtime-specific preparation and the free space needed by the rest of your workstation.

Compatibility belongs to the runtime, too

The card points to Strata for its consumer-GPU serving path. It also dates its llama.cpp compatibility statement to October 23; that date matters because an architecture-support claim can change as implementations develop. A GGUF extension alone does not establish that an arbitrary installed runtime can execute this architecture.

Strata’s versioned DeepSeek setup guide is the place to check the specific build, additional downloads and memory/storage conditions. Treat support for a model in one engine as a deployment path to investigate, rather than permission to reuse the configuration of a different model.

This is a quantized derivative of an existing checkpoint

The quantization release builds on DeepSeek’s upstream 1017 checkpoint. The quantized repository identifies the base checkpoint’s MIT license. Credit for the base model and credit for ISTA-DASLab’s quantization work should both remain visible.

Our next useful question is practical: what does a complete deployment cost in storage, RAM, startup time and task latency on the machine available? Published benchmark comparisons can guide that investigation, but they do not answer it for our hardware. We have deliberately kept a local trial separate from this source-reviewed release coverage.

Primary sources: ISTA-DASLab model card and size table, upstream DeepSeek model card, and the linked versioned Strata setup guide.