Hey HN! Author here.
Built this because deploying ML models is painful. Python + pip dependencies + Docker = 5GB+ images. For edge/embedded/air-gapped systems, this doesn’t work.
LOOM loads HuggingFace transformers directly in Go. No Python runtime. ~10MB binary.
Technical highlights:
– Native safetensors parser
– Pure Go BPE tokenizer (no transformers library)
– Full transformer stack (MHA, GQA, RMSNorm, SwiGLU)
– Cross-platform determinism (MAE < 1e-8)
– Published to PyPI, npm, NuGet
Tradeoff: CPU-only, 1-3 tok/s on small models. Correctness first, speed second.
Works with Qwen, Llama, Mistral, SmolLM. Cross-compiles everywhere Go runs.
Demo: https://youtu.be/86tUjFWow60
What layer types should I add next? Currently have: Dense, Conv2D, MHA, RNN, LSTM, LayerNorm, RMSNorm, SwiGLU, Softmax (10 variants), Residual.
Questions welcome!
Comments URL: https://news.ycombinator.com/item?id=45874310
Points: 1
# Comments: 0
Source: github.com
