An Ollama alternative that installs nothing
Ollama made local LLMs easy by wrapping llama.cpp in a daemon, a CLI and a model registry. It is very good, and it needs a machine you are allowed to install software on. This playground covers the other case: open a URL, pick SmolLM2 or Qwen3, and the weights download into the browser cache and run on your GPU through WebGPU. No binary, no background service, no port, no admin rights — which is the difference between "I will try this later" and trying it on a locked-down work laptop, a Chromebook or someone else’s machine.
At a glance
- Install
- None — it is a web page
- Model format
- ONNX from Hugging Face (not GGUF)
- Runtime
- WebGPU via transformers.js and WebLLM, WebAssembly fallback
- Model ceiling
- Roughly 0.5B–3B parameters
- Local HTTP API
- No — Ollama has one, this does not
- Admin rights needed
- No
- Where Ollama wins
- 7B+ models, an OpenAI-compatible endpoint, a far larger catalog
Where Ollama is clearly better: model size. It will happily run 7B, 14B and larger quantised models against your full system RAM and VRAM, and that is a different quality tier from anything a browser can hold. It also gives you an OpenAI-compatible local HTTP endpoint, model files you own on disk, and a much larger catalog. If you want a local coding assistant, you want Ollama or LM Studio, not a browser tab.
Where the browser is better: distribution and isolation. There is nothing to install or update, the model cache is scoped to the origin and evictable rather than a directory you have to remember to clean up, and a link is enough to hand the whole thing to someone else. It is also the only option when installing software is not permitted — which, in most companies, is the actual constraint.
The practical ceiling here is roughly 0.5B to 3B parameters, because WebGPU exposes a per-binding buffer limit rather than your VRAM and a browser tab has a memory budget of its own. SmolLM2-135M loads in seconds on almost anything; SmolLM2-1.7B and Qwen3-0.6B are noticeably better and want a discrete GPU. Expect roughly 20–60 tokens per second for a 0.5B model on mid-range discrete graphics.
ollama alternativerun llm without installingollama in browserlocal llm no installllama.cpp browser alternativelm studio alternativebrowser llm chatllm on chromebookrun llm without admin rightswebgpu llm
Frequently asked questions
Can this run the same models as Ollama?
Not the same files. Ollama runs GGUF through llama.cpp; a browser runs ONNX through ONNX Runtime Web or MLC-compiled weights through WebLLM. Several model families exist in both — Qwen, Llama, Phi, SmolLM — but you need the ONNX or MLC build, not the GGUF one.
How much worse is a browser-sized model?
Meaningfully worse at facts and multi-step reasoning, close enough for rewriting, summarising, extraction, classification and short Q&A. A 1.7B model is not a 7B model and we would rather say that plainly than let you find out on a task that matters.
What about LM Studio or Jan?
Same trade-off as Ollama: a desktop application with a bigger model ceiling, a real GUI and a local server, in exchange for an install and an OS you control. Pick this when you cannot or would rather not install anything, or when you want to hand someone a link.
Where are the weights stored, and can I delete them?
In the browser Cache Storage API under the key "transformers-cache". The storage page lists what is cached and deletes models individually, and clearing site data removes all of it.
Is my data uploaded to a server?
No. The model weights are downloaded from Hugging Face to your browser once, and every inference after that runs on your own GPU through WebGPU. Your text, images and audio are never sent anywhere — hjLabs.in has no inference server and no way to see your input.