UDGAM LABS · MLX · 8-bit, 161.4 MB

Laghu-MLX-8bit

Laghu, a 152M-parameter English chat model trained from scratch by Udgam Labs for $146 of pretraining compute, converted to MLX (8-bit) for Apple silicon Macs. Recommended for Macs: half the size of the bf16 build at almost the same quality; about 456 tokens per second on a Mac mini (Apple M6).

Try it: in your browser, on your device · code on GitHub · udgamlabs.com/models/laghu

Quick start

pip install mlx-lm
mlx_lm.generate --model udgamlabs/Laghu-MLX-8bit --prompt "Who are you?" \
  --chat-template-config '{"enable_thinking": false}'

For an OpenAI-compatible server: mlx_lm.server --model udgamlabs/Laghu-MLX-8bit --chat-template-args '{"enable_thinking":false}'. The MLX server turns thinking on by default; the switch above turns it off. Sampling defaults: temperature 0.7, top-p 0.95, top-k 64. Keep context at 16,384 tokens or less.

Results

Measured by us on 2026-10-22 on a Mac mini (Apple M6, 24 GB), mlx-lm 0.32.0:

File	Size	Perplexity	Mean KL vs F16	Top-1 agreement	Generation speed
Laghu-MLX-8bit (8-bit)	161.4 MB	52.37	0.0040	96.2%	about 456 tokens/s

Held-out text: 32 windows of 2,048 tokens, compared with the F16 GGUF reference. MLX applies Laghu's 512-token sliding window, and the chat format (thinking off and on, multi-turn, tool calls) was checked in the MLX server. We do not publish a 4-bit MLX build: it loses clearly (mean KL 0.337, top-1 agreement 69.9%).

For the model's own results, see the Laghu model card.

Where it falls short

Laghu is a small research model: weak at maths (GSM8K 7.9% with thinking) and instruction following (IFEval 26.1), knows little, and can loop with greedy decoding. Details on the model card.
The MLX server turns thinking on by default; use the switch above for normal chat.

How it was trained

See the Laghu model card and the technical report. Converted from the released safetensors with mlx_lm.convert (mlx-lm 0.32.0).

Intended use and limitations

Local, offline use and research on Macs. Not for factual answers or decisions that matter; light safety training.

Licence

Citation

@misc{udgamlabs2026laghu,
  title  = {Laghu: a fully open 152M chat model trained from scratch},
  author = {Prasad, Jitesh K},
  year   = {2026},
  howpublished = {Udgam Labs},
  url    = {https://huggingface.co/udgamlabs/Laghu}
}

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com