MicroLLM Lab Lets You Compare Small AI Models in the Browser, No Install Required
Download a model and put it through conversations and benchmarks on your own device. The project's title says seven models, but the official repository lists eight, including GPT-2.
Developer robss2020 has released MicroLLM Lab, a web tool that runs small language models directly in the browser. Users can pick a model, chat with it, and compare benchmark results across models on the same screen. It is essentially a web-based laboratory for checking how a small AI answers, and how fast, on your own computer, without setting up a development environment.
How it works in the browser
MicroLLM Lab downloads model files and generates responses on the user’s own device. It relies on WebGPU, the technology that lets a browser tap into a device’s graphics processing power. The developer says the model files have been compressed to a 4-bit (Q4) format, and that downloaded files are kept in the browser’s internal storage.
The official repository provides links to a web demo and a full download of the files. According to the developer, there is no need for an API key for an external AI service, and no Python or CUDA installation is required. What you do need is a browser and device that support WebGPU. After choosing a model, you can either start a conversation or run a prepared set of benchmark questions.
Seven models, or eight?
The number of models in the project’s headline description does not match the number listed in the repository. The title touts seven models, while the official repository list names eight, including GPT-2. The models on the list range from 26 million to 362 million parameters. Compressed to Q4, file sizes run from 15MB to 216MB depending on the model, so the size of the download varies with your choice.
The developer explains that the existing PetitGPT model and its architecture were ported to WebGPU, with comparison features added on top. The result goes beyond simply running a single model: it lets you see the speed and accuracy of several models in one place. The list includes both SmolLM, from 2024, and its successor SmolLM2, so the differences between the two generations can be compared as well.
What about speed and answer quality?
In measurements the developer took on an Apple M4 with Safari, PetitGPT generated text faster than SmolLM2 135M Instruct, at 115 tokens per second versus 66. MicroLLM Lab shows benchmark accuracy alongside speed as a comparison metric.
There is a limit to how much a conversation can handle at once. This build caps context at 2,048 tokens. The developer notes that small models can get arithmetic wrong or produce plausible-sounding falsehoods. On languages, the developer says MiniMind2 is stronger in Chinese than in English.
MicroLLM Lab is a tool for testing firsthand how quickly and accurately a small AI answers on your own device. Its main use is throwing the same question at several models and weighing the results side by side.