garagelm
AI & ML interests
Frontier-Compression
Recent Activity
GarageLM
Chasing the most intelligent model on the cheapest hardware.
Every model here was pretrained, fine-tuned, evaluated, and benchmarked on a single Apple M4 Pro Mac mini — no cluster, no cloud, ~$3 of electricity per flagship run. The research axis: how much measured capability can one consumer machine produce?
Models
| Model | What it is |
|---|---|
| hybrid-gpt-232m | 232M base · hybrid local+global attention · 1B refined tokens |
| hybrid-gpt-232m-chat | SmolTalk SFT of the base — zero benchmark regression |
Headline result: at matched, locally re-run evaluation the base model is at-or-above Pythia-160M (trained on 300× more tokens) across HellaSwag / PIQA / ARC-Easy / WinoGrande within ±2.8-pt slice resolution, and ahead of GPT-2 on ARC-Easy — the one gap beyond 2×SE — with ~5× fewer training FLOPs (6ND, one-epoch-WebText assumed). The hybrid attention runs at ~30% of full attention's KV cache and decodes on-device at ~310 tok/s fp16 (530+ at 4-bit, MLX).
Quick start
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("garagelm/hybrid-gpt-232m-chat")
model = AutoModelForCausalLM.from_pretrained(
"garagelm/hybrid-gpt-232m-chat", trust_remote_code=True)
prompt = "User: Why is the sky blue?\nAssistant:"
out = model.generate(**tok(prompt, return_tensors="pt"),
max_new_tokens=200, eos_token_id=50256)
print(tok.decode(out[0]))
These are honest research artifacts, not products: fluent, format-reliable, and factually unreliable in the way every ~200M model is — limitations are stated prominently on each card.
Small hardware. Real research.