As of this moment (4:37 p.m. EST / 1:37 p.m. Pacific on February 2, 2026), I've set Claude Code to port a significant piece of software from its C++ source code to webGPU. Nobody's done it before.
The software is a significant one: Microsoft trained a highly efficient model on millions of books, and the model itself is super fast and efficient, and only takes about 400 MB of RAM, while being efficient enough to run on CPU. It's able to engage in conversation, follow instructions, and do some simple coding tasks.
In the environment I've set up, Claude is in a VM and has been instructed to plough through any difficulties, do its own research, do anything needed to get the model working in webGPU.
Claude Code has performed more significant ports, to far more obscure languages and environments than WebGPU, but this one is significant due to the importance of the efficient model that it is porting.
If Claude Code succeeds at its porting task, which it is expected to do overnight, then it will result in a major, highly efficient model that is able to run entirely in the browser at moderate cost.
Note: in an ironic twist, the environment Claude is in doesn't actually have a GPU, so I've asked it to just try really, really hard to make sure it works on webGPU as well as processor. So it's flying blind and has no way to test its own code except by reasoning about it.
As of this moment (4:37 p.m. EST / 1:37 p.m. Pacific on February 2, 2026), I've set Claude Code to port a significant piece of software from its C++ source code to webGPU. Nobody's done it before.
The software is a significant one: Microsoft trained a highly efficient model on millions of books, and the model itself is super fast and efficient, and only takes about 400 MB of RAM, while being efficient enough to run on CPU. It's able to engage in conversation, follow instructions, and do some simple coding tasks.
In the environment I've set up, Claude is in a VM and has been instructed to plough through any difficulties, do its own research, do anything needed to get the model working in webGPU.
Claude Code has performed more significant ports, to far more obscure languages and environments than WebGPU, but this one is significant due to the importance of the efficient model that it is porting.
The model is licensed under the MIT license.
Here is ChatGPT's description of the model that it is porting: https://chatgpt.com/share/698116d7-d008-800c-a295-aaf2cfc7a8...
And here is the model itself: https://huggingface.co/microsoft/bitnet-b1.58-2B-4T/tree/mai...
And its weights: https://huggingface.co/microsoft/bitnet-b1.58-2B-4T?
The model is available via the web at: https://bitnet-demo.azurewebsites.net/
If Claude Code succeeds at its porting task, which it is expected to do overnight, then it will result in a major, highly efficient model that is able to run entirely in the browser at moderate cost.
Note: in an ironic twist, the environment Claude is in doesn't actually have a GPU, so I've asked it to just try really, really hard to make sure it works on webGPU as well as processor. So it's flying blind and has no way to test its own code except by reasoning about it.
God speed, little coder.