Niko1221/Strata

What it does

Strata is a local inference stack for running large quantized models on consumer hardware, including NVIDIA RTX 20–50 series, recent AMD Radeon cards, Intel Arc on Linux, Apple Silicon, and Ryzen AI Max Strix Halo (41). It ships a desktop chat app and a terminal launcher, and exposes OpenAI- and Anthropic-compatible APIs on localhost so coding tools such as Claude Code and Codex CLI can use the local model (49).

It supports several model families: Qwen3.8-Flash-Next (including an IQ1_M coder variant), Qwen3.6-35B-A3B, and DeepSeekV4.1-Flash-1017. Larger models are read partly from the SSD, so disk space matters (21). The README provides memory and token-speed measurements for specific hardware, but these are author-reported figures for particular versions, not guarantees for every machine.

Install

The simplest path is the desktop app: download the installer for Windows, Linux, or Mac from the latest release (4), install it, choose a model, and start chatting. On first start, the app downloads the inference engine for your GPU (28).

If you prefer a terminal, install with the one-liner and then run strata launch. On Linux or macOS:

curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh && sh install.sh
strata launch

On Windows, use PowerShell:

irm https://github.com/Niko1221/Strata/releases/latest/download/install.ps1 | iex

After installation on Windows, open a new terminal and run strata launch. The command prompts you to download and start a model (31). Then use strata connect to print connection settings for your preferred coding tool (50). Building from source, Docker, and full install steps are documented separately (38).

Constraints

You need compatible hardware and enough VRAM, system RAM, and in some cases a large amount of free disk space for the larger model packs. For example, the DeepSeekV4.1-Flash-1017 options need roughly 197–280 GB depending on the quantization set (21). Model quality and speed vary widely with quantization, GPU, and driver stack; the README’s benchmark table covers specific hardware and versions only.

The desktop app is distributed as prebuilt installers and the terminal path uses installer scripts. The repository describes the project as MIT-licensed (64), but the README does not map a particular release binary back to repository source, so the build should be treated as a release artifact you choose to trust, not as something verified by the documentation. Image input is advertised but not documented in depth in the README; model-specific guides should be checked before relying on vision tasks.

HotRepo's take

Strata’s Hacker News discussion in October 2026 was substantial: a thread with 940 points and 420 comments drew both positive and negative reports. One commenter reported 124 tokens/s on an RTX 4090 with 128 GB DDR5 and a Ryzen 7950X3D (1). Other commenters raised quality concerns, including a vision benchmark where Strata produced a much larger coordinate error than llama.cpp with the same GGUF and vision adapter weights, plus an anecdote about the model losing track of files during coding (1). Those reports are individual anecdotes, not controlled evaluations, and another commenter noted that vision results can depend on how the vision tower is quantized (1). The practical takeaway is that Strata is relevant for people who want to run large models on their own machines and can test quality against their own workloads; the release-binary distribution and the open questions about output quality are the main things to weigh before adopting it.