Strata

Overview

Strata is an open-source local inference engine and desktop app for running Qwen3.8-Flash-Next, a 125-billion-parameter model, on consumer hardware. It provides one-click installation for Windows and Linux, a terminal client, OpenAI- and Anthropic-compatible APIs on localhost, optional image input, and chat and monitoring tools while keeping chats and files on the PC.

Key Features

• Runs Qwen3.8-Flash-Next, a 125-billion-parameter model, on consumer PCs with at least 32 GB of combined RAM and graphics memory.
• One-click app installers for Windows and Linux, plus a macOS app and a one-line terminal installer.
• OpenAI-, Anthropic-, and OpenAI Responses-compatible APIs on localhost for connecting coding agents and other apps.
• Optional picture input, with the image reader running on the CPU or GPU depending on available graphics memory.
• Desktop chat with model, GPU, CPU, and RAM monitoring, a Models page, Settings, web search, and optional multi-model support.
• Terminal commands for launching, monitoring, updating, diagnosing, and connecting Strata to tools such as Claude Code, Codex CLI, and OpenCode.

Use Cases

• Run a large Qwen model locally on a gaming PC without relying on a cloud API.
• Connect coding agents and other OpenAI- or Anthropic-compatible apps to a local model endpoint.
• Chat with a model that can read pictures while keeping chats and files on the PC.
• Monitor local inference speed, GPU, memory, and requests with the desktop app or terminal.

Getting Started

• Download the Strata app installer for your system and graphics card from the releases page, then run it; Windows installs per user without administrator rights.
• On Linux, make the AppImage runnable with chmod +x and install FUSE 2; on Mac, open the DMG and drag Strata to Applications.
• Alternatively, run the one-line install script for Linux or macOS, or the PowerShell installer for Windows, then run strata launch.
• Follow the guided prompts to choose a model, context size, picture support, and access, then download the engine and model.

Deployment & Requirements

• At least 32 GB of RAM and VRAM combined; 64 GB of RAM runs every model size.
• About 80 GB of free disk; an SSD makes the first start much faster.
• Windows 10 or 11, Linux, or macOS, plus a current graphics driver.
• Supported hardware includes NVIDIA RTX 20/30/40/50 series with 12 GB or more, AMD Radeon RX 6800 and newer, Intel Arc, Apple Silicon Macs with macOS 15 or newer, and AMD Ryzen AI Max unified-memory systems.
• The model download is about 40-70 GB and resumes if interrupted.

Before You Adopt

• License: MIT. Review its terms before using, modifying, or distributing the project.
• The Windows installer is not code-signed, so Windows may show a protection warning.
• Web search is an explicit exception to local-only operation because search terms leave the PC.
• Initial model loading can make the PC slow or unresponsive for one to three minutes.
• The RX 6000 engine is built but not yet tested on real RX 6000 hardware.
• Gyro model support on AMD is experimental.