CaaLM

/

CaaLM-v1

Like 1

Follow

Code as a Language Model 4

Hi @ereniko ! Great question. CaaLM-v1 ("Code as a Language Model") is a 1.5B parameter model that predicts what a piece of code would print — without actually running it. No compiler, no runtime, no interpreter.

A few key points:

• Base: fine-tuned from Qwen2.5-1.5B (Qwen2 architecture, Apache-2.0 license)
• Input format: you give it code after Code: and it completes the Output: section with the predicted stdout
• The twist: it wasn't trained just on real languages. It learned from Python, JavaScript, Lua, COBOL plus 200 synthetically generated fake languages with randomized syntax but consistent semantics — so it learned the idea of execution rather than one language's syntax
• Result: it can handle languages it's never seen, scoring ~96% overall on the benchmark, including 100% on Python, JS, Lua, and invented languages like SCRIBBLE @x BECOMES 7 / YELL @x + @y

If you want to try it hands-on, there's a demo Space here: CaaLM-v1-Demo. The README has more examples and details on the benchmark and known limitations. Happy to answer follow-ups!

@VelocityBot does this model has quantizations?

I'll check the Hub for any quantized versions of this model.