CaaLM/CaaLM-v1
TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kPublished:Apr 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

CaaLM-v1 by CaaLM is a 1.5 billion parameter model based on Qwen2.5-1.5B, designed to predict the output of code without a compiler or interpreter. It was uniquely trained on a mix of real and 200 synthetically generated programming languages with randomized syntax but consistent semantics, enabling it to infer execution logic for languages it has never encountered. This model excels at understanding abstract execution concepts and is primarily used for predicting stdout for basic programs, including those in novel, unseen syntaxes.

Loading preview...

CaaLM-v1: Code Output Prediction for Any Language

CaaLM-v1 (Code as a Language Model) is a 1.5 billion parameter model built on Qwen2.5-1.5B, specifically engineered to predict the output of code. Unlike traditional compilers or interpreters, CaaLM-v1 infers program execution without needing a predefined language specification.

Key Capabilities

  • Language-Agnostic Execution Prediction: Trained on a diverse dataset including Python, JavaScript, Lua, COBOL, and 200 synthetically generated languages with randomized syntax, CaaLM-v1 can predict outputs for programming languages it has never explicitly seen before.
  • Abstract Execution Understanding: The model learns the concept of execution rather than specific language rules, allowing it to infer semantics from novel syntaxes.
  • High Accuracy on Basic Operations: Achieves 96.2% overall accuracy on benchmarks, including 100% on Python, JavaScript, Lua, and novel fake languages for assignment, printing, and loops.
  • Handles Core Programming Constructs: Reliably processes variable assignment, arithmetic, print statements, conditionals (if/else), and while loops with accumulator patterns.

Good For

  • Understanding Code Execution Logic: Ideal for research into how LLMs can abstractly learn program execution.
  • Predicting Outputs of Simple Programs: Effective for determining the stdout of short, basic code snippets across various languages, including those with unfamiliar syntax.
  • Educational Tools: Can be used to demonstrate fundamental programming concepts and execution flow without needing a full runtime environment.
  • Rapid Prototyping/Testing: Quickly get an expected output for simple code without compilation, especially useful for novel or domain-specific languages.