Description
Running a model that can write code and read images on your own machine usually dies on two things: not enough VRAM, and an afternoon of environment setup. A 125-billion-parameter model is server territory, so nobody expects a gaming PC to carry it.
Strata turns that into download, open, press install. It runs Qwen3.8-Flash-Next and starts at 32 GB of RAM and VRAM combined. The engine decides which parts live on the GPU and which in system memory, and falls back to a low-RAM mode by itself when memory is tight. On an RTX 5070 with 64 GB of RAM the author measured the smallest Q2_0 build at 94 tokens per second for answers and 2,650 for reading your prompt.
Once installed it opens an OpenAI- and Anthropic-compatible API on localhost, so Claude Code, Codex CLI and OpenCode connect straight away. Chats and files stay on your PC, it is MIT licensed, and there is no account and no telemetry.
One-click installers: Separate downloads for Windows, Linux and Mac, with builds for NVIDIA, AMD, Intel Arc and Apple Silicon. The model download is 40 to 70 GB and resumes if interrupted.
Model picked for your memory: The app recommends a size from your memory: Gyro-S and a code-focused Coder from 32 GB, IQ3_S at 64 GB, and closer-to-original quantizations above that.
Local compatible API: OpenAI- and Anthropic-style endpoints on your machine, so coding agents and apps plug in unchanged. An MCP server lets AI tools install, start and stop Strata themselves.
Image input: On by default, so you can hand screenshots and pictures to the model.
Terminal mode: `strata launch` asks about model, context, images and who may connect, then serves. `strata monitor` shows speed, GPU, memory and requests live. From 0.2.4 the Terminal also runs DeepSeek V4.1-Flash.
Built-in chat and workspaces: The desktop app has chat, a monitor and a models page. Chat can work inside a folder you pick and, in Manual mode, asks before editing a file or running a command. Web search sends only the query out.
Strata turns that into download, open, press install. It runs Qwen3.8-Flash-Next and starts at 32 GB of RAM and VRAM combined. The engine decides which parts live on the GPU and which in system memory, and falls back to a low-RAM mode by itself when memory is tight. On an RTX 5070 with 64 GB of RAM the author measured the smallest Q2_0 build at 94 tokens per second for answers and 2,650 for reading your prompt.
Once installed it opens an OpenAI- and Anthropic-compatible API on localhost, so Claude Code, Codex CLI and OpenCode connect straight away. Chats and files stay on your PC, it is MIT licensed, and there is no account and no telemetry.
Features
One-click installers: Separate downloads for Windows, Linux and Mac, with builds for NVIDIA, AMD, Intel Arc and Apple Silicon. The model download is 40 to 70 GB and resumes if interrupted.
Model picked for your memory: The app recommends a size from your memory: Gyro-S and a code-focused Coder from 32 GB, IQ3_S at 64 GB, and closer-to-original quantizations above that.
Local compatible API: OpenAI- and Anthropic-style endpoints on your machine, so coding agents and apps plug in unchanged. An MCP server lets AI tools install, start and stop Strata themselves.
Image input: On by default, so you can hand screenshots and pictures to the model.
Terminal mode: `strata launch` asks about model, context, images and who may connect, then serves. `strata monitor` shows speed, GPU, memory and requests live. From 0.2.4 the Terminal also runs DeepSeek V4.1-Flash.
Built-in chat and workspaces: The desktop app has chat, a monitor and a models page. Chat can work inside a folder you pick and, in Manual mode, asks before editing a file or running a command. Web search sends only the query out.
