Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM/Strata Review: A 125B-Parameter Open Model on a Gaming PC
Overview & Reference

Strata Review: A 125B-Parameter Open Model on a Gaming PC

·9 min read·By Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI
Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

Strata (version 0.2.2, released 21 October 2026) is a free, MIT-licensed app and Terminal tool that, per its README, runs one 125-billion-parameter open model on an ordinary gaming PC and serves it to your other programs through a local API. Every speed, quality and privacy statement in this review comes from the project; PromptQuorum has not installed or tested it.

Strata is an MIT-licensed engine, desktop app and Terminal tool that, per its README, runs Qwen3.8-Flash-Next, a 125-billion-parameter open model, on a gaming PC by sharing the work between graphics memory, RAM and the SSD. This review reflects the project's README, safety notes and v0.2.2 release page as checked on 23 October 2026.

Strata Review: A 125B-Parameter Open Model on a Gaming PC

Key Takeaways

  • Maintenance: the repository was created on 24 September 2026 and versions 0.2.0-beta, 0.2.1 and 0.2.2 appeared between 19 and 21 October 2026, so it is active but only a month old; no stale-project warning applies.
  • License: MIT for Strata itself; each model file carries its own licence.
  • Cost: free, with no account; support is voluntary.
  • Installers: the Windows installer is not code-signed yet, per the README.
  • Basis: the README, safety notes and release page, read on 23 October 2026, with no hands-on testing.

📍 In One Sentence

Strata is an MIT-licensed engine and desktop app that runs a single 125-billion-parameter open model on a gaming PC.

💬 In Plain Terms

It spreads the model across your graphics card, RAM and SSD, then answers in a chat window or through a local API that other tools can call.

📌Note: PromptQuorum has not installed, tested or benchmarked Strata. Speed, quality and privacy statements below are the project's own.

What Is Strata?

Strata is an inference engine, a desktop app and a Terminal client built around one model, Qwen3.8-Flash-Next, which the README describes as a 125-billion-parameter model that normally needs a server. The Qwen team publishes the model on Hugging Face; Strata, published under the GitHub account Niko1221, runs it on consumer hardware.

The README explains the trick with a library: the model is a team of 24,576 small specialists ("experts") and each word needs only 10 of them. The graphics card holds the most-used experts, RAM holds all of them and the SSD holds a 29 GB lookup table. A small helper also guesses the next few words, which the big model checks in one pass; the project says this returns the same answer 1.6 to 1.8 times sooner. The engine uses parts of llama.cpp, and the code lives at Niko1221/Strata.

Get It

Strata is downloaded from the Releases page on GitHub, one installer per system and graphics card.

System and card
Download
Windows, NVIDIA (driver 580+)windows-cuda13.exe
Windows, older NVIDIA or driverwindows-cuda12.exe
Windows, AMD RX 9000 / R9700windows-hip.exe
Windows, AMD RX 7000 / Strix Halowindows-hip-rdna3.exe
Linux, NVIDIA (driver 580+)linux-nvidia-cuda13.AppImage
Linux, older NVIDIA or driverlinux-nvidia-cuda12.AppImage
Linux, AMD RX 9000 / R9700linux-amd-rx9070.AppImage
Linux, AMD RX 7000linux-amd-rx7900.AppImage
Linux, Strix Halolinux-amd-strixhalo.AppImage
Mac, Apple Silicon (macOS 15+)mac.dmg
Terminal, Linux and Macinstall.sh (one line from the README)
Terminal, Windowsinstall.ps1 (one line from the README)
Source codegithub.com/Niko1221/Strata
Safety and privacy notesdocs/SAFETY.md

Version reviewed: 0.2.2 (engine 0.1.47), published on 21 October 2026 per the GitHub release page, which is the latest upstream update at the time of checking. No separate product website was found: the repository's homepage field and the maintainer's profile link are empty, and a web search turned up only the repository and third-party write-ups. The README says the Windows installer is not code-signed, so Windows may show "Windows protected your PC", and Linux AppImages need FUSE 2. RX 6000 cards have no installer of their own, and the README says that engine is not yet tested on a real RX 6000 card.

What You Need

Strata needs at least 32 GB of RAM and graphics memory (VRAM) combined, about 80 GB of free disk, a current graphics driver and Windows 10/11, Linux or macOS 15 or newer. On a Mac or a Ryzen AI Max, the unified memory counts for both.

Computer
Support per the README
NVIDIAGeForce RTX 20, 30, 40, 50 series, 12 GB or more
AMDRadeon RX 6800 and newer, Radeon AI PRO R9700, 12 GB or more
IntelArc cards; the smallest model is tested on an Arc A750 (Linux) only
Apple SiliconM1 or newer, macOS 15 or newer; Intel Macs cannot run it
Ryzen AI Max (Strix Halo)Unified-memory systems, own installers for Windows and Linux

Per the README, two or three cards of the same make can share a model; Strata never splits a model across NVIDIA and AMD cards. Older cards and processors have separate experimental pages in the repository.

Pricing

Strata is free: the README states an MIT license, no account and no cloud.

Item
Cost (USD)
App, Terminal and engine$0 (MIT license)
Model filesNo price listed; own licences apply
Support the projectOptional, via Buy Me a Coffee

Read from the README on 23 October 2026. It also invites sponsors to make contact through the maintainer's GitHub profile. The model download is the real cost: about 40 to 70 GB per the README.

How to Get Started

The steps follow the README's "Option A: the Strata app"; PromptQuorum has not run them.

  1. 1
    Compare your PC with the requirements
    Why it matters: Check memory, free disk and graphics card against the table above before downloading.
  2. 2
    Download the installer for your system and card
    Why it matters: Pick the matching file from the Get It table. Terminal users can run the one-line script instead and then start Strata with `strata launch`.
  3. 3
    Install and let the app read your PC
    Why it matters: The first screen recommends a model that fits; press install. The model download is large and continues where it stopped if interrupted.
  4. 4
    Wait through the first load
    Why it matters: The README says the PC can be slow or stop responding for 1 to 3 minutes while the model loads into RAM, longest the first time. Do not close the window.
  5. 5
    Chat, or connect your other tools
    Why it matters: Use the app's Chat page, or run `strata connect` to print settings for OpenCode, Codex CLI, Claude Code and others.

Features Confirmed by the Sources

Every item below comes from the README or the project's docs.

  • Local API. The server offers an OpenAI-compatible API, an Anthropic-style API on the same port and the OpenAI Responses API for Codex CLI; strata connect prints settings for OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent and OpenClaw.
  • Pictures. Picture input is on by default for models that can read it; the reader runs on the processor when the graphics card has under 24 GB.
  • Chat app. The app has a Chat, a Monitor of model, GPU, CPU and RAM, a Models page and Settings; in Manual mode the chat asks before it changes a file or runs a command.
  • Several cards. Each graphics card can be set to Main, Helper or Off, and a thinking level (off, low, medium or high) can be chosen per chat.
  • MCP. AI tools can install, start and stop Strata through its MCP server.
  • Terminal. strata launch asks for model, context, pictures and access one question at a time; flags let scripts and AI agents run it without prompts.

Model Sizes by Memory

Strata ships one model in several compressed sizes; the README says smaller sizes are faster and larger ones a bit smarter.

Memory (RAM + VRAM)
Size suggested by the README
32 GBGyro-S, or the Coder for code
40 GBSwift 1.5 IQ2_XS, or Gyro-M
48 GBIQ3_XXS: slower, better
64 GBIQ3_S: slowest, best
96 GB or moreUnsloth UD-IQ4_XS up to Q8_0

The README counts RAM plus VRAM less 5 GB that the system keeps. Gyro-S needs 27.6 GB of that memory for a 58.5 GB download and, per the README, has lower quality than IQ2_XS; Gyro-M needs 35 GB for a 92 GB download. The README says IQ3_S matches the full model on the published tests; PromptQuorum has not checked those tests.

Speed: The Project's Numbers

The README reports 94 tokens per second for its fastest size and 53 for IQ3_S on an RTX 5070; PromptQuorum has not reproduced any figure. A token is about three quarters of a word.

Size and card
Answers (tok/s)
Prompt (tok/s)
Q2_0, RTX 5070 (12 GB)942,650
IQ3_S, RTX 5070 (12 GB)531,620
Coder, RTX 5070 (12 GB)552,180
Q2_0, RX 9070 XT (16 GB)601,160

Source: the README's "How fast is it?" tables, measured by the project on a Ryzen 5 7600 with 64 GB RAM (RTX 5070) and a Ryzen 9 3900X with 47 GB RAM (RX 9070 XT). The README adds that Strata reads the first message of a chat in full, at about one minute per 30,000 tokens, and that later messages start in seconds.

Privacy and Network Use

Strata's safety notes say chats, prompts, files, pictures and models stay on your PC, that there is no telemetry and no account, and that starting or running a model never downloads anything. They add that, once a model is downloaded, Strata runs without internet, and that the server listens on your PC only unless you give other devices access with an API key.

Per the same notes, the internet is used only when you ask: model, engine and app-update downloads from Hugging Face and GitHub Releases, the agent's web search (DuckDuckGo, with Bing as the second source) and web pages, outside model services you add, and a Cloudflare tunnel if you switch it on. Web search sends the search words off your PC.

📌Note: The notes say they come from reading the code of this version, but they are the maintainer's statements. PromptQuorum has not audited the code or the network traffic.

Who Should Use It

  • Owners of a recent gaming PC or a high-memory Mac who want one very large open model without renting a server.
  • Developers who use Claude Code, Codex CLI or OpenCode and want a local OpenAI- or Anthropic-style endpoint.
  • People who want chats kept on their own machine, per the project's safety notes.

What We Could Not Verify

  • Hands-on quality and speed. PromptQuorum did not run Strata, so answer quality, the speed figures and the claim of matching the full model are the project's.
  • Privacy and download checks. The no-telemetry, offline and signed-update statements were not checked against the code or network traffic.
  • Maturity. The repository dates from 24 September 2026 and showed 697 open issues on 23 October 2026; the Mac and Intel Arc paths are new in 0.2.2, and the project names few tested machines for them.
  • Model choice. The pages checked describe one model in several sizes plus a few named alternatives, not a general model library.

Competitors and Alternatives

llama.cpp

Platforms:
Windows, macOS, Linux
Price / license:
Free / MIT
Key difference:
The C++ engine behind many tools, including parts of Strata; command line first
Articles about llama.cpp (9)

+159 more not shown

Lucebox

Platforms:
Linux
Price / license:
Free / Apache-2.0
Key difference:
Hand-tuned kernels and speculative decoding for specific consumer GPUs
Articles about Lucebox (2)

Competitor details change often; confirm each app's current price, license and platforms on its own listing.

Compare in directory

Frequently Asked Questions

Does Strata work without a graphics card?

The README says the CUDA, HIP and SYCL engines need a card to start, so a PC needs a supported one. The Mac engine can run without a separate card.

Can I use Strata with Claude Code?

Per the README, yes, through the Anthropic-style API by setting ANTHROPIC_BASE_URL to the local address. The model that answers is Qwen3.8-Flash-Next, not a Claude model.

Can I reach it from my phone or another PC?

The README names strata launch --access lan, which creates an API key by default; --no-api-key skips the key and prints a warning.

Is the Windows installer safe to run?

The README says it is not code-signed yet and says the Terminal installer checks a SHA-256 hash. PromptQuorum has not verified either statement.

Verdict

Strata is a free, MIT-licensed app and Terminal tool that, per its README, makes a 125-billion-parameter open model usable on a gaming PC and exposes it through OpenAI- and Anthropic-style APIs; it shipped three releases in the week before this check and installs by double-click on Windows, Linux and Apple Silicon Macs. Against that, every speed, quality and privacy statement comes from the project, the repository is about a month old, the Windows installer is not code-signed and it is built around one model rather than a library. It suits people with enough combined memory who want that model for chat or coding agents; readers who want many models and a longer track record can compare Ollama or llama.cpp.

Sources

← Back to Power Local LLM

MCP