ranausmanai / tinyforge

What is tinyforge?

Tinyforge lets you run a 0.8B parameter language model on your laptop that self-improves at coding tasks without cloud, teachers, or human feedback. It generates solutions for problems like fizzbuzz or roman numerals, tests them automatically, evolves better ones via failure-driven search, and fine-tunes itself with LoRA on repair pairs—all in Python using MLX on Apple Silicon with just 6GB RAM. Fire up the CLI with `tinyforge --model models/mlx-q4-qwen35-08b --quick` for a 10-minute demo showing baseline to post-training gains.

Why is it gaining traction?

Unlike bloated cloud APIs or pre-trained giants, tinyforge proves self-play works for code on tiny models, jumping single-pass scores from 46% to 92% via evolution and repair training. The hook is the local, verifiable loop: generate, test, mutate, learn—echoing github tiny recursive model experiments but production-ready with sandboxed eval and custom tasks. Devs dig the no-data-leak privacy and hardware efficiency, akin to miniforge setups.

Who should use this?

ML engineers prototyping self-improvement on edge devices, indie devs building local code assistants without API costs, or researchers extending to SQL, Verilog, or math proofs where outputs are verifiable. Perfect for Apple Silicon users tweaking tiny models like those in tiny models gallery, avoiding github tiny windows 11 bloat.

Verdict

Grab it for experiments—solid docs, reproducible results, quick CLI make the 1.0% credibility score (15 stars) forgivable for such an early, innovative technique. Maturity lags (limited benchmarks), but the self-training loop scales to bigger models; run the demo before committing.

(198 words)