thalyn-labs/SmolLM2-360M-4bit

SmolLM2-360M, 4-bit, for MLX

A 4-bit MLX conversion of HuggingFaceTB/SmolLM2-360M, the base model (no instruction tuning), published by Thalyn Labs for Kaleidokin, a game about what a language model is, learned by raising one. It is the game's second kind of newborn: the English-trained, smaller one. The weights are exactly the upstream weights, quantized; nothing was trained.

• Architecture: Llama (model_type: llama), 362M parameters, 49,152-token vocabulary, tied embeddings.
• Quantization: affine, 4 bits, group size 64 (4.50 bits per weight), 204 MB of safetensors.
• Made with mlx-lm on 2026-09-06:

mlx_lm.convert --hf-path HuggingFaceTB/SmolLM2-360M -q --q-bits 4 --q-group-size 64 --mlx-path SmolLM2-360M-4bit

Loads with mlx-lm (Python) and mlx-swift-lm (Swift) like any mlx-community model:

mlx_lm.generate --model thalyn-labs/SmolLM2-360M-4bit --prompt "The light through the window was" --ignore-chat-template

Licence: Apache 2.0, as upstream. Cite SmolLM2 (arXiv:2502.02737) for the model itself.