Contents
Strata is an open-source tool that runs local LLMs on ordinary Windows, Linux, and Mac PCs and exposes OpenAI- and Anthropic-compatible APIs on localhost. Setup is straightforward, but the published speeds come from the developer’s measurements, so you won’t know whether your PC can match them until you try.
Key Points#
- It works through both a desktop app and a terminal, reportedly without requiring an account.
- Its local API is compatible with OpenAI and Anthropic, allowing tools such as Claude Code and Codex CLI to connect to it.
- Large DeepSeek-family models read whatever cannot fit in memory from the SSD, requiring more than 250GB of free disk space.
- The speed table was measured with Strata 0.2.5; it does not say whether conditions are the same for the currently distributed 0.2.6.
- The performance of “Qwen3.8-Flash-Next” has not been independently verified beyond this material.
What It Can Do#
At its core are an engine that loads models and runs inference, and a mechanism that exposes it as an API on localhost. The desktop version combines chat, files, and live previews in one screen. Terminal users can do the same with strata launch.
Named supported tools include Claude Code, Codex CLI, OpenCode, and Hermes Agent. Connection instructions are shown by strata connect. Still, API compatibility and practical usability with each tool are separate matters; the latter requires individual testing.
Getting Started#
On Linux and Mac, use the following commands to install Strata and launch a model. On Windows, run the official install.ps1, open a new terminal, and run strata launch.
curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh
sh install.sh
strata launchOn first launch, the app downloads an additional engine suited to your graphics card. The ISTA-DASLab versions of models are recommended, while Unsloth versions and Qwen3.6-35B-A3B are also available. Before deciding which will fit on your PC, check the storage and memory requirements in the page’s comparison table.
Reading the Numbers and Caveats#
The published measurements show output speeds of 4–6 tokens per second on an RTX 5070 (12GB, 64GB RAM) and 9–11 tokens per second on an RTX 3060 (12GB, 46GB RAM). Both use 2.00 bpw versions of DeepSeek-family models. Output speeds around 4 tokens per second are quite slow for an interactive conversation as you read.
There are three caveats. Measurements were taken with Strata 0.2.5, with no guarantee of matching results on the current 0.2.6. The developer performed the measurements, and conditions such as prompt length and temperature are not documented in detail. And the 2.00 bpw version alone is about 257GB, requiring roughly 265GB of free space.
Licenses also need to be considered separately. Strata itself uses the MIT license, but models have their own licenses. Read each model’s license before commercial use.
Next Step#
Check your PC’s GPU model, RAM capacity, and free disk space against the figures above. If free disk space falls below the candidate model’s requirements, rule out the DeepSeek family, select Qwen3.6-35B-A3B through strata launch, and try giving it one of your everyday coding tasks.
Source: https://github.com/Niko1221/Strata
Contents
