The prevailing assumption in language model development is that capability is bought with scale. That assumption holds at the frontier and holds poorly everywhere else.
Vex-Amber-Fable-2.0 is a 2-billion-parameter causal language model built to test how much software engineering capability can be recovered at a size that runs on hardware most organisations already own. It uses float32 precision and an 8k context window.
Results
On SWE-bench (Verified) the model reaches 65.37%. On HumanEval it reaches 60.98%, placing it alongside models in the 8B to 30B range. On LiveCodeBench, which is harder to contaminate, it reaches 44.19%.
What we think is actually going on
The gains come from training data fidelity and architectural discipline rather than from any single trick. Benchmarks are also imperfect instruments, and we would rather publish the evaluation code than ask anyone to take the numbers on faith.
Weights, datasets and evaluation harness are available on Hugging Face.