Strata's README puts the instruction I would read first in a warning box, not in the speed table. While the model starts, your PC can be slow or stop responding for one to three minutes, because Strata loads 35 to 55 GB into RAM and locks part of it for the graphics card. Strata runs Qwen3.8-Flash-Next, a 125-billion-parameter model, on supported NVIDIA and AMD gaming PCs with at least 12 GB of VRAM and 32 GB of system RAM, and much of the model sits in that RAM.
The project calls itself a one-click install for Windows and Linux, with experimental community paths for Apple Silicon, Intel Arc, AMD Strix Halo and older GPUs. It serves the model on localhost with OpenAI- and Anthropic-compatible endpoints and optional image input. The README says nothing leaves your PC.
The "Which model should I pick?" table is organized by RAM, with one exception: on a 32 GB system, a 24 GB card also runs Q2_0 and IQ2_XS. The requirements ask for 12 GB of VRAM or more and 32 GB of RAM or more, and the table maps memory to quantization:
The README lists what you give up at the small end. The Coder version reaches 91% of the full model's SWE-bench Verified score as measured by its authors, and it is weaker outside code, including Chinese and other CJK text. For those, the README points you back to sizes that keep every expert. The larger Unsloth files work with less RAM, but Strata then reads part of them from the SSD while it answers, which the README calls much slower: 7 to 8.5 tokens per second for UD-Q4_K_XL on a 64 GB PC.
The README's speed tables come from two machines it describes as ordinary gaming PCs: an RTX 5070 with 12 GB, a Ryzen 5 7600 and 64 GB of RAM, and an RX 9070 XT with 16 GB, a Ryzen 9 3900X and 47 GB. On the NVIDIA box it reports 94 tokens per second of output at Q2_0 and 53 at IQ3_S, with prompt reading between 1,620 and 2,650 tokens per second on a 32K-token input. The engine versions differ between rows, and the full tables are in docs/DETAILS.md.
For agent use, the README also covers concurrency and first-message time. By default Strata answers one request at a time and queues the rest; setting "parallel": 2 allows two, and on a 12 GB card that makes each answer slower. And Strata reads the first message of a chat in full at about one minute per 30,000 tokens, while follow-ups start in seconds. If your agent opens with a large context dump, I would expect to wait on that first message. The BENCH.bat script (./bench.sh on Linux) measures chat, 4K and 16K prompts and an agent-style session using your own settings.
Once running, Strata opens a browser app at http://127.0.0.1:8080 with Chat, a live Monitor of the model and your hardware, and an About page. For other tools you add an OpenAI-compatible provider at http://127.0.0.1:8080/v1, and the README says any API key and any model name work. Anthropic-style clients use /v1/messages, and for Claude Code the README sets ANTHROPIC_BASE_URL=http://127.0.0.1:8080, and the Responses API for Codex CLI lives at /v1/responses.
The install itself can be handed to an agent: the README offers a one-line prompt asking your coding assistant to follow docs/AI_SETUP.md. The assistant then checks your card, RAM and disk and picks a model. There is also an MCP server so AI tools can install, start and stop Strata.
Serving to your phone or another PC uses START-HERE.bat --setup --host 0.0.0.0 --api-key <secret>, and the README adds: always set a key. If you only need the model on the same machine, I would leave the host setting alone.
For a first run, budget about 80 GB of free disk, preferably SSD, and a roughly 70 GB download that resumes if interrupted. Expect the slow start the README warns about, and run the benchmark before you decide which size to keep.
**GitHub:** [https://github.com/Niko1221/Strata](https://github.com/Niko1221/Strata)
*Curated by [Agent Palisade](https://www.agentpalisade.com) — practical AI for small and mid-sized businesses.*