Scores
Table with columns: probe, Tin 2B, Qwen3.5 2B| probe | Tin 2B | Qwen3.5 2B |
|---|
| Says it is Tin | 93.8 | 56.2 |
| Tin knowledge | 63.1 | 22.7 |
| Writes Tin | 54.1 | 5 |
| Fixes Tin | 92.5 | 63.6 |
| Fixes stacked bugs | 90.2 | 24.4 |
| Agent loop · native | 55 | 6.2 |
% pass on a held-out set, graded by the real Tin compiler. "Writes Tin" means the program compiles and prints exactly the expected output on the first try.
Run it
curl -fsSL https://tin.ad/install.sh | bashtin model pull tin:2b # then any agent without a model: line uses it
Or directly with mlx-lm:
pip install -U mlx-lmmlx_lm.generate --model cagataydev/tin-2b-v3-4bit --prompt "Write a Tin program with a guarded tool fn deleteFile."
All sizes
Table with columns: size, writes Tin, fixes Tin, agent loop, phone build| size | writes Tin | fixes Tin | agent loop | phone build |
|---|
| 0.8B | 37.5 | 88.8 | 32.5 | tin-0.8b-4bit |
| 2B | 54.1 | 92.5 | 55 | tin-2b-v3-4bit |
| 4B | 68.3 | 97.2 | 65 | tin-4b-4bit |
Limits
- About 46% of first tries fail
tin check or print the wrong thing, so always run the compiler.
- It knows Tin as of the training commit; newer std functions may be missing.
- Small model: details like flag names and error codes are often wrong.
Licence
Apache-2.0, like the base Qwen3.5 model.