On this page

Strata is a free, open-source local runner. Its v0.2.5 release (GitHub published_at 2026-10-25T10:06:04Z, about 18:06 Beijing time on October 25, 2026) puts DeepSeek V4.1 Flash into the desktop app on Windows and Linux—not a claim that DeepSeek itself shipped a new official desktop product. The real floor is not “any 12GB card”: the setup still needs tens of gigabytes of system RAM and a multi-hundred-gigabyte download.

Disk and RAM to plan for

Release notes for 0.2.4 introduced DeepSeek V4.1-Flash in the Terminal and published the hardware framing that 0.2.5 still uses for sizing:

  • GPU: about 12GB VRAM or more.
  • RAM: about 46GB worked on the authors’ RTX 3060 box; 64GB is the tested comfortable size. More free RAM means a larger expert store and faster answers.
  • Download: about 239GB (IQ1_M), 257GB (IQ2_XXS), or 274GB (IQ2_XS). That full Strata download includes the Engram table and the DSpark drafter—so it is larger than the backbone-plus-Engram totals on the Hugging Face model card alone.

DeepSeek V4.1-Flash is a sparse mixture-of-experts model. Strata keeps the experts it uses most in a RAM store and streams the rest from SSD with direct reads—so the whole model does not have to fit in VRAM. In 0.2.5, that RAM store can hold every expert when free RAM is enough (it used to stop at 48GB) and still leaves about 12GB free on the machines they describe. That is an upgrade to how much it can cache, not a rewrite of the “plan for roughly 46–64GB RAM” floor.

App vs Terminal, and which platforms get DeepSeek

In 0.2.5, both the app and the Terminal are at 0.2.5. The model picker lists the three official sizes with a plain note on size, speed, and disk that warns and never blocks. Context is fixed at 8K; there is no calibration step. A Speed choice (Faster or Full quality) is new: Faster uses 5 of the model’s 6 experts per token and is the default on PCs with 16GB of graphics memory or less.

The DeepSeek engine is a separate package (strata-dsv41) beside the Qwen/Metal engines. On the v0.2.5 assets page, the dsv41 packages are Windows and Linux only (CUDA and one HIP build). Mac has a 0.2.5 dmg and a small Metal engine inside, but no dsv41 and no DeepSeek on this release.

Quality claim: three sizes, three benchmarks

The weights come from ISTA-DASLab’s GSQ-RCO GGUF builds of DeepSeek-V4.1-Flash-1017—not a DeepSeek-branded “new model drop.” On the model card, the IQ2_XXS (2.00 bpw) build matches the base model on AIME25, GPQA-Diamond, and LiveCodeBench v6 (100.0% of that three-benchmark task average at reasoning effort 50). IQ1_M keeps about 98.7% of the same average. That is a claim on those three scores under the card’s settings, not a blanket “2-bit equals lossless everywhere.” Strata’s own 0.2.4 notes also publish perplexity and code/agent text deltas for each size; IQ2_XS is the pick when disk allows and coding/agent text closeness matters most.

For the inference stack DeepSeek itself open-sourced around V4.1 Flash, see DeepInfer. Strata is a different, third-party path: installer + local app/Terminal, not DeepInfer itself.

Speed figures (for reference only)

From the 0.2.4 notes, on the 2.00-bit set the authors report roughly 5 tokens/s output on an RTX 5070 12GB with 64GB RAM (code), and about 8–10 tokens/s on an RTX 3060 12GB with 46GB RAM. Prompt read rates at 8K are also listed there. In 0.2.5, they add that Faster mode decoded code and agent runs about 16% faster on an RTX 5070, with smaller gains on other boxes, plus several engine tweaks (CPU starting on already-loaded experts while the rest read from SSD, fewer GPU launches per layer, and so on). Drive speed matters a lot on 12GB cards. All of these numbers are project-reported—for reference only.

How this sits next to other DeepSeek notes

The Flash 1017 open-weight release and later rankings are covered in How DeepSeek V4.1 Flash 1017 Ranks on Code Arena, Vals AI and Artificial Analysis Benchmarks. The open inference framework behind the official side is covered in DeepSeek Open-Sources DeepInfer, the Inference Framework Behind V4.1 Flash. For local coding harnesses rather than local weight runners, see Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does and DeepSeek Harness Install Guide: Windows, Linux, CLI, Plugins.

Whether those tokens/s and three-benchmark recoveries hold on a given card, RAM, and SSD is untested here—for reference only. Feedback can go through the site-owner contact link in the footer.