Strata
A free, open-source app that runs very large AI models on an ordinary gaming PC, with no account and no cloud, and plugs into coding agents like Claude Code.
01 · Introduction
What Strata is
Strata is a free, open-source program (MIT license) that lets you run big AI models on your own computer, including machines with an ordinary graphics card. Chats, files and pictures stay on your PC, and no account is needed.
The problem it solves: the strongest open models are huge. Normally a model must fit in the graphics card's memory (VRAM), and a gaming card has only 8 to 24 GB. Strata gets around this with models built as a mixture of experts (MoE). Such a model is split into many small "experts", and only a few of them work on each word. So Strata keeps the experts used most often in VRAM, the rest in your normal memory (RAM), and for the very largest models reads what is left from your SSD. You trade some speed for being able to run a model that would otherwise not start at all.
What you get:
- A desktop app (Windows, Linux, Mac) with chat, files and a live preview, and a terminal app (
strata launch). - Models it can download and run for you, such as Qwen3.8-Flash-Next, Qwen3.6-35B-A3B and DeepSeek V4.1-Flash. The project recommends models in compressed ("quantized") form, which are smaller than the original at some cost in quality.
- A local server that speaks the OpenAI and Anthropic API formats, so coding agents (Claude Code, Codex CLI, OpenCode and others) can use your local model.
New words: a local model runs on your machine instead of a company's server. Quantization stores the model's numbers with fewer bits so the file is smaller. Tokens per second is how fast the model writes its answer.
Name warning: other projects are also called "Strata". This handbook is about the local-model engine at github.com/Niko1221/Strata.
02 · When to use it
When it is useful
- Privacy. Your code, documents and pictures never leave your computer. Once a model is downloaded, Strata works without internet.
- No per-token bill. After the download, running the model costs only electricity.
- A local model for your coding agent. Point Claude Code, Codex CLI or similar tools at it for tasks you do not want to send to the cloud.
- Trying large open models without renting a server.
- Learning how local AI works on hardware you already own.
Be realistic: a big model squeezed onto a gaming PC is slower than a cloud service, and a heavily compressed model can be a little worse than the full one. If you need top speed or the strongest closed models, a cloud API is still the better tool.
03 · Setup
How to get started
Check your computer first. Supported: NVIDIA GeForce RTX 20/30/40/50 series, AMD Radeon RX 7000/9000 series and AI PRO R9700, Intel Arc (Linux), Apple Silicon Macs (M1 or newer, macOS 15+), AMD Ryzen AI Max, and NVIDIA DGX Spark. Windows 10/11 and Linux x86-64 are supported; Intel Macs are not.
How much memory? What matters is RAM plus VRAM. The project's guide lists, for example, about 38 to 55 GB combined for the recommended Qwen3.8-Flash-Next packs (a 12 GB graphics card with 32 to 64 GB of RAM is a typical gaming PC), and a smaller 35B model at 32 to 52 GB of RAM. Leave memory for Windows and your other apps. DeepSeek V4.1-Flash is far bigger and needs about 245 to 280 GB of free disk, because part of it is read from the SSD.
Install, the easy way:
- Download the installer for your system from the latest release (Windows setup, Linux AppImage, or Mac dmg).
- Open the app and choose a model. On first start it downloads the engine for your graphics card.
- Start chatting.
Or use the terminal:
# Linux / Mac
curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh && sh install.sh
# Windows PowerShell
irm https://github.com/Niko1221/Strata/releases/latest/download/install.ps1 | iex
Then run strata launch and follow the prompts. To see what is missing before downloading anything, add --plan.
These commands download and run software from the internet, so read the install script first if you want to be sure, and only download from the official GitHub release page.
04 · How to use it
Day-to-day use
Chat in the desktop app. Pick a model, type, and it answers. You can also give it files and pictures (image input works with some models).
Connect a coding agent. Strata runs a local server that mimics the OpenAI and Anthropic APIs. Run:
strata connect
It prints setup steps for your tool. The project lists Claude Code, Codex CLI, OpenCode, oh-my-pi, pi.dev, Hermes Agent and OpenClaw, and others work through the same two APIs.
Watch it run. The terminal app shows model activity, speed, memory use and recent requests, which helps you see whether the model fits well.
Pick a different size. Larger packs answer better but are slower and need more memory. The project's default recommendation is ISTA-DASLab's packs: Q2_0 is the fastest, IQ3_S the best quality. If a model is slow or crashes, step down one size.
Speed. Speeds depend heavily on your hardware. For the huge DeepSeek V4.1-Flash model, the author's own measurements show about 4 to 6 tokens per second on an RTX 5070 with 64 GB RAM, and about 9 to 11 on an RTX 3060 with 46 GB RAM. Those are the author's numbers, so test on your own machine.
05 · Pitfalls
Things to watch out for
- It is new and changes fast. The project is only a few months old and ships updates every few days (version 0.2.11 on 5 November 2026). Commands, model names and sizes may change, so check the README.
- Heavy on memory and disk. Models are tens to hundreds of GB. Check free disk and the memory table before you download.
- Slower than the cloud. Large models on a gaming PC are noticeably slower than hosted services, and a very compressed pack can lose some quality.
- Speed claims vary. Articles about Strata quote different speeds and requirements, and some mix up graphics cards. Trust the project's own docs and your own test.
- Model licenses differ. Strata is MIT-licensed, but each model has its own license. Check it before using a model commercially.
- Installers run code. The install commands download and run programs. Use the official GitHub release page only. Strata says its downloads are checked against fixed hashes and that it sends no telemetry.
- Keep the server private. By default it listens on your PC only. If you open it to other devices or the internet, use the API key and understand the risk.
- A coding agent can change your files. In its default mode the built-in agent asks before it changes anything. Keep it that way until you trust it.
06 · Resources