AhoosAI Studio 2.5: the model does the work
Connect a folder and the model writes the files and runs the commands itself, asking you before every change. Also new: internet access if you want it, a battle between the base model and its adapter, and any model from Hugging Face.
Until this version, asking a small model to "make hello.py and run it" got you a code
block to copy — and once, the invented output of a run that never happened. Told about its
tools in words, it kept answering in prose. So the choice is no longer left to prose: each
step of the model's turn is one decision, held to a grammar that admits a tool with its
arguments or the answer, and nothing else. On Nimbus 1.1 at its smallest size, on a
four-core laptop, the same request now ends with the file on disk, hello world
from running it, and one sentence saying so.
The model is also told where it is: that it runs inside the Studio on your computer, which folder it may touch, which shell your system uses and whether the internet is on. Reading is free. Every write, delete and command is shown to you first — the change, or the command exactly as it will run — and nothing happens until you allow it.
Three more things are new. Internet access, off until you choose: your computer's own connection with no key, an API key for Tavily, Brave, Exa, Serper or Jina, or any search service you describe. Battle: one prompt answered side by side by the base model and by the same model with its adapter, together or one after the other, with a blind vote if you want it. And any model from Hugging Face: paste a link to a model and to an adapter, and the files are found and queued for download; an adapter in PEFT form is converted on your computer.
What it costs is time. On the processor, a first message takes about a minute before the answer starts, and the steps after it are quicker. Pictures are not supported in the new page yet. Builds are ready for Windows, macOS (Apple silicon and Intel) and Linux.