Qwen3.5 0.8B: the model for a browser you know nothing about
470MB of q4f16 weights, 512MB of GPU buffer, and code that compiles. Why Qwen3.5 0.8B opened our on-device catalogue.
Qwen3.5 0.8B is the model we give to a browser we know nothing about. It is
small, around 470MB once quantised in q4f16, and it still writes code that
compiles and produces tool calls that parse. For a first model in the
catalogue, this was the right compromise.
What you really download
The parameter count does not interest us very much. What interests us is the
number of bytes that must cross the network before the first token arrives.
Here it is ~470MB, taken from the Hugging Face Hub as
onnx-community/Qwen3.5-0.8B-Text-ONNX.
The browser keeps the weights in cache, so the second visit costs nothing. The first one costs what it costs, and no engineering on our side will change this. It is the price of running the model on your machine instead of ours. We think it is a fair price.
The memory
| Budget | Value |
|---|---|
| Working set | ~800MB |
| Minimum GPU buffer | 512MB |
| Precision | q4f16 |
The number to verify on your own hardware is the 512MB buffer. WebGPU exposes
maxBufferSize and maxStorageBufferBindingSize in the adapter limits, and
the integrated GPUs of a few years ago are often below what one imagines. Our
AI Readiness Test reads these limits and tells you
what will fit, before you spend the download for nothing.
The sampling
Qwen publishes its own decoding preset and we do not try to be clever:
{ "temperature": 0.7, "top_p": 0.8, "top_k": 20 }
A model of 0.8B is much more sensitive to sampling than a model of 7B. If you push the temperature above 0.9, you obtain text that reads well and which is false. For a small model this is the worst possible failure, because nothing in the form of the answer warns you. So we stay conservative.
Where it is useful
Qwen3.5 0.8B passes through the ordinary text-generation path of
Transformers.js. No custom runtime, no patch for the family. This is precisely
why we keep it: when something breaks, it is the model that tells us whether
the problem comes from the model or from our own plumbing.
For chat, for a summary, for a short edit in a file, it does the work. For a whole repository, no. In that case look at Qwen3.5 2B, or Qwen3.5 4B if your GPU allows it.
It is in the Local LLM workspace, and what you type there stays in your tab.