Strata

Niko1221/StrataReviewed by AI

What it does

Strata runs large open-weight models, such as Qwen3.8-Flash-Next and DeepSeek V4.1-Flash, on a normal computer with a gaming graphics card. There is one installer per system: you pick a model, the app downloads the engine that matches your card and the model files from Hugging Face, and then you chat in its own window. It also serves an OpenAI- and Anthropic-compatible address on your machine, so Claude Code, Codex CLI, OpenCode and similar tools can use the local model instead of a paid service. No account is needed, but the downloads are large.

How easy it is to install

B · installer plus one simple step · Windows

Windows B
macOS B
Linux B

Windows is the easiest: one Strata-0.2.9-windows-x64-setup.exe, then you choose a model and the app downloads the engine for your graphics card on first start. macOS has a .dmg (Apple Silicon M1 or newer only) and Linux an AppImage you make executable. None is an A because nothing works until a model is downloaded, and your graphics card has to be on the supported list (NVIDIA RTX 20-50, several AMD Radeon models, Intel Arc on Linux, Apple Silicon).

Numbers on GitHub

Stars: 37,090on Oct 31, 2026
Downloads: 1,842,284857,538 this week
Growth: +4,250stars in 6 days

Stars over time

Facts

Runs on: Windows, macOS, Linux
You need to install: Graphics card
To build from source: Python, Docker, NVIDIA graphics card
Disk space: 107 MB–322 MB · from the installer
Accounts or keys: None
Cost: Free · MIT licence, no account and no fees; the models are free downloads. The README asks for voluntary support through Buy Me a Coffee. The real cost is hardware, electricity and disk space.
License: MIT (Permissive)
Can replace: ChatGPT Plus, Copilot, OpenRouter
Last update: Oct 31, 2026

Risk signals (7)

AttentionVery new repository with many stars
37095 stars in 37 days
automatic check
AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-qfhq-4f3w-5fph, GHSA-rrmf-rvhw-rf47, PYSEC-2026-139; transformers: GHSA-29pf-2h5f-8g72, GHSA-69w3-r845-3855, GHSA-fgcw-684q-jj6r; pillow: GHSA-22hg-r758-rrw9, GHSA-24rc-8hj2-v98c, GHSA-254f-3936-288h
automatic check
Attentiondisk and download far larger than the installer
The installer is 107-278 MB, but the engine packages in the same release reach 892 MB and the model files are much bigger: the README's DeepSeek table lists 239 GB, 257 GB and 274 GB sets needing "about 245 GB" to "about 280 GB" of free disk. The Qwen setup also fetches a draft layer of about 5 GB.
AI review
Attentionterminal install runs a downloaded script
The terminal install line is `irm https://github.com/Niko1221/Strata/releases/latest/download/install.ps1 | iex` on Windows, and on Linux and Mac `curl -fsSL .../install.sh -o install.sh && sh install.sh`: a script is fetched and run straight away. The desktop installers avoid this.
AI review
Infomodel files come from third party accounts
setup.py pins each Hugging Face repository to a fixed commit with a size and SHA-256 for every file (ISTA-DASLab, unsloth, ukisai, agentionai). It also documents an opt-out, STRATA_ALLOW_UNPINNED=1, which takes the file from the repository's current version "without the SHA-256 check", and HF_ENDPOINT can point the download at a mirror.
AI review
Infoopens a local api that can be exposed
The model is served on http://127.0.0.1:8080 with an OpenAI- and Anthropic-compatible API. setup.py documents `--host 0.0.0.0 --api-key KEY` to reach it from other machines, which puts the model on your network.
AI review
Infomodel licences are not the project licence
The program is MIT, but the README links separate "Credits and model licenses", and the Qwen and DeepSeek weights it downloads carry their own terms, which can restrict commercial use.
AI review

Evidence is shown in its original language.

What Radar could not check

Whether it really runs at a usable speed on your exact graphics card and amount of memory: the speed table in the README was measured by the project itself.
How long the first run takes, between the engine download, tens or hundreds of gigabytes of model files and the preparation step.
Whether the installers are signed, and what Windows or macOS warn about when you open them.
Whether the answers from these heavily compressed models are good enough for your work.
Whether the pinned dependency versions with known vulnerabilities (torch, transformers, pillow) are used by the parts you run, since torch and transformers are only installed for one optional conversion step.