The Pulse
Strata Runs Qwen’s 125B Model on Gaming PCs, Reports Selected Speedups
Strata says its inference engine runs Qwen3.8-Flash-Next on consumer PCs by distributing work across the GPU, system memory and processor. Release v0.1.43.1, published October 12, fixes issues in the previous release, while v0.1.43 reports
AI.info Team · Written with AI ·
“That is why it generates fast on a gaming PC where a dense 27B is slow, and also why it does not match the 27B on hard coding tasks: it is a wider, faster model, not a bigger dense one.”
Niko1221, Strata repository owner, in a comment on GitHub issue #1422, October 8, 2026
Niko1221 made the comparison in a response to a user discussing Qwen3.8-Flash-Next and a dense 27-billion-parameter model. Strata’s repository describes the engine as a way to run the 125-billion-parameter Qwen model on Windows or Linux PCs. The project says the model can chat, write code, read pictures and connect to applications and coding agents through local APIs.
How Strata runs the model on a PC
Strata’s explanation says the model has 24,576 experts, with each token using 10. The project describes distributing work across the PC: the graphics card keeps frequently used experts, system RAM holds the full set, and the processor works on experts that are not on the GPU. The SSD stores a lookup table. Strata also describes a helper model that guesses upcoming tokens for the larger model to check, and says this approach produces the same answer 1.6 to 1.8 times sooner.
The project’s listed baseline is a supported NVIDIA or AMD graphics card with at least 12 GB of VRAM, 32 GB of RAM and about 80 GB of free disk space. It recommends using an SSD. The installer downloads the model, about 70 GB, and opens a local browser interface. Strata warns that the first startup can make a PC slow or temporarily unresponsive while it loads 35 to 55 GB into RAM. Users can connect compatible apps through OpenAI-compatible or Anthropic API endpoints on the local machine.
Published speeds and their limits
The repository reports 94 tokens per second for Q2_0 on an NVIDIA RTX 5070 with 12 GB of VRAM, a Ryzen 5 7600 and 64 GB of RAM. For an AMD RX 9070 XT with 16 GB of VRAM, a Ryzen 9 3900X and 47 GB of RAM, it lists 60 tokens per second for Q2_0. These are the project’s own measurements on two PCs, not independent benchmark results. The repository says the NVIDIA Q2_0 result used engine 0.1.36; other NVIDIA rows used 0.1.26. It does not give an engine version for the listed AMD results.
The releases page gives a separate comparison for v0.1.43 against v0.1.42. It describes the figures as the project’s interleaved tests, generally using five pairs and reporting medians. Results vary by hardware and model size. For example, the notes list IQ3_XXS decoding gains of 21% to 38% on a Radeon AI PRO R9700 using a 16 GB card’s cache, and 14% to 21% for IQ3_S on an Intel Arc Pro B70. By contrast, the notes say Q2_0 and IQ3_XXS decoding on an RTX 5070 running Windows were unchanged by default.
The latest release is a hotfix
Strata’s releases page lists v0.1.43.1 as the latest release, published October 12, 2026. The release describes it as a hotfix for v0.1.43, with fixes including restoring Linux NVIDIA decode performance on some desktops, addressing long-prompt stalls on Windows and SATA SSDs, and improving two-card batch serving. It says the answers remain the same as in v0.1.43 in the project’s identity tests, with a stated exception for user messages containing a particular tool-call string.
The v0.1.43 release, published October 11, reported performance changes across selected NVIDIA, AMD and Intel configurations. Its notes also say the Windows RTX 5070’s default Q2_0 decode performance was unchanged in that release. Strata’s repository says the engine is open source under the MIT license, while some components and all model files have their own licenses. The project offers a benchmark tool that runs a set of tests on a user’s own installation in about five to ten minutes.