TinkyBrain 31M — A Language Model Built From Scratch
31 million parameters. Trained from zero. On a Mac mini M4.
No fine-tuning. No pre-trained weights. No shortcuts. Every single weight in this model was learned from scratch on Apple Silicon using MLX.
What Makes This Special
Most "custom models" are LoRA adapters slapped on top of someone else's foundation model. TinkyBrain is different:
- Custom architecture — MicroBrain transformer built from scratch in
model.py
- Custom tokenizer — grown organically during curriculum training, not borrowed from GPT/LLaMA
- Custom training pipeline — DJ playlist-style curriculum learning (70 tracks, added one at a time like ingredients in a recipe)
- Trained entirely on Apple Silicon — MLX on M4 Mac mini, no cloud GPU needed
- Purpose-built for AAC — designed to power speech prosthetics for non-verbal users (stroke survivors, neurodivergent individuals)
Training Approach
The playlist trainer (train_playlist.py) treats training data like a DJ set:
- Start with baby-level language (greetings, feelings, basic needs)
- Add one "track" at a time — 100 epochs per track on the cumulative dataset
- Each track builds on what came before — the model never forgets early lessons
- 14 curriculum stages from babbling to full conversation
- Smart early stopping with patience-based checkpointing
70 JSONL training files covering: greetings, feelings, play, school, family, friends, animals, imagination, food, health, safety, identity, culture, and more.
Files
checkpoints/v3_best.safetensors — Best v3 model (71MB)checkpoints/best.safetensors — Best v1 model (80MB)checkpoints/full_best.safetensors — Full curriculum model (36MB)checkpoints/curriculum_step_*.safetensors — Every curriculum checkpointmodel.py — MicroBrain architecturetokenizer.py — Custom tokenizerchat.py — Inference + generation
Usage
1import mlx.core as mx
2from model import MicroBrain
3from chat import SimpleTokenizer, generate_greedy
4
5# Load model + tokenizer
6tokenizer = SimpleTokenizer.from_file("checkpoints/v3_tokenizer.json")
7model = MicroBrain(vocab_size=len(tokenizer), d_model=512, n_heads=8, n_layers=8)
8model.load_weights("checkpoints/v3_best.safetensors")
9
10# Generate
11response = generate_greedy(model, tokenizer, "How are you feeling?")
12print(response)
Built With
- MLX — Apple's machine learning framework for Apple Silicon
- TinkyOven — Custom macOS SwiftUI app for visual playlist-style training
- TinkyBrain — Chrome extension for harvesting training data from the web
Part of the Tinky Ecosystem
TinkyBrain powers the AAC speech prosthetics in:
- TinkySpeak — Android AAC device for non-verbal users
- TinkyAAC — iOS/Android speech prosthetic for stroke survivors
- TinkyTown — Accessibility kiosk for municipal buildings
- TinkyAsk — Universal business ordering system
Built by a dad who wanted nothing more then his non verbal son and his mom who who had a stroke able to speak.
No cloud GPUs were harmed in the making of this model.
