zimengxiong/talkie-1930-13b-it-mlx

Talkie 1930 13B IT - MLX

MLX-compatible BF16 safetensors for talkie-lm/talkie-1930-13b-it.

This repo keeps the original Talkie tensor names/layouts and includes a lightweight mlx_talkie runtime for the custom architecture.

Run with bundled runtime

hf download zimengxiong/talkie-1930-13b-it-mlx --local-dir talkie-1930-13b-it-mlx
cd talkie-1930-13b-it-mlx
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m mlx_talkie.cli --model-dir . --max-tokens 80 "Write a short note about radio."

Run with mlx-lm

Talkie support for mlx-lm is proposed in https://github.com/ml-explore/mlx-lm/pull/1220 . With that branch installed:

python -m mlx_lm.generate \
  --model . \
  --prompt "<|user|>Write a short note about radio.<|end|><|assistant|>" \
  --max-tokens 80

Notes

• Architecture: 40-layer, 40-head decoder-only GPT, RoPE base 1,000,000, SwiGLU MLP, RMS normalisation, embedding skip connections, per-head/per-layer gain parameters.
• Precision: BF16.
• Context length: 2048 tokens.
• Source model: https://huggingface.co/talkie-lm/talkie-1930-13b-it
• Source code: https://github.com/talkie-lm/talkie