GitHub radar
Strata runs a 125B model on your own gaming PC, no cloud
Strata is a free installer and local server that runs the Qwen3.8-Flash-Next model on a gaming PC or Apple Silicon Mac, and exposes it through an OpenAI- and Anthropic-compatible API.
Strata is an installer and local server for the Qwen3.8-Flash-Next model (125 billion parameters), which is normally built for server hardware. The first-run wizard checks your GPU and memory, asks which model version and how much context you want, then downloads a ready-made engine build for your card (NVIDIA, AMD or Intel) or compiles it on the spot if none fits. After install, the model runs as a local web server on 127.0.0.1:8080 that speaks the same protocols as the OpenAI and Anthropic APIs, so Claude Code, Codex CLI, OpenCode and other tools can connect to it by pointing at a different server address. The model reads images by default and works with files and code. The project is MIT-licensed, has no telemetry and needs no account; the repository has a separate document describing exactly what it downloads and when it touches the network.
Why a vibe-coder should care
This is for people who want to run a big model offline and for free, without a subscription and without sending their files to someone else's cloud. By task class I would not put this setup next to what I actually use myself: Qwen is a Chinese model, and for real agent work it is noticeably behind Opus and Fable in Claude Code. The strong side here is not model quality, it is how carefully the install and hardware matching are done: releases are signed, and the install script checks the checksum of what it downloads before running it, which is rare for a tool that fetches and builds binaries. I would use this for simple repeating tasks and offline experiments, not as a replacement for a paid agent.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install Strata from https://github.com/Niko1221/Strata — download the installer for my system from the releases page, go through the setup wizard, start the local server, point Codex CLI or OpenCode at it as an OpenAI-compatible endpoint, and show that the model answers a simple question. No model key is needed, but ask me if anything else needs downloading.
You need 32 GB of combined RAM and VRAM plus about 80 GB of free disk: an Apple Silicon Mac (unified memory) or a gaming PC with a 12 GB+ NVIDIA/AMD card works — a regular laptop without a modern chip or GPU will not run it.
Open on GitHub▌ More finds