On October 25, 2026, OpenClaw DC said a GSQ-RCO build of DeepSeek V4.1-Flash can run locally on a 12GB NVIDIA GPU when loaded through the Strata runtime.
In the OpenClaw DC write-up, the model is described as a 763.2B-parameter MoE, with reported output speed around 4–6 tokens per second on an RTX 5070. That lower VRAM figure comes with larger system requirements elsewhere: roughly 46–64GB of RAM, at least 239GB of free disk, and NVMe storage handling part of the model data.
The main limitation is software support. OpenClaw says these GGUF files are not a standard llama.cpp drop-in today because the package uses different metadata and keeps an Engram component separate, so Strata is the runtime used for this setup. Reported support is limited to x86_64 Windows and Linux systems with compatible NVIDIA RTX 30, 40, or 50 series GPUs; Macs are not listed.
For teams evaluating local inference, the bigger story is the tradeoff behind the 12GB headline: lower VRAM, but substantial RAM, storage, and runtime constraints. DeepSeek’s transparency page does not describe this quantized release, so the performance and compatibility details should still be treated as OpenClaw-reported rather than broadly confirmed.