# Lattice Quark 1.5B (MLX 4-bit)

MLX conversion of the Quark 1.5B nanochat-style language model, trained from scratch. Optimized for Apple Silicon: transformer linears quantized to 4-bit (group size 64), token and value embeddings kept in bf16.

- Parameters: 1.5B (total, incl. value embeddings)
- Layers: 26, hidden 1536, 12 heads, head dim 128
- Context: 2048 tokens, RoPE base 100k
- Vocab: 32768 (GPT-4-style BPE)
- Model type: `nanochat2` (complete architecture: value embeddings, smear, backout, residual lambdas)
- Size: ~1.7GB

## Files

- `model.safetensors` — weights (4-bit quantized linears + bf16 embeddings)
- `config.json` — model configuration (`model_type: nanochat2`)
- `tokenizer.json` / `tokenizer_config.json` — HuggingFace tokenizer (byte-exact with the training tokenizer, incl. chat template)
- `nanochat_tokenizer.json` — original tokenizer dump
- `nanochat_rope.safetensors` — precomputed rotary embeddings
- `generation_config.json` — eos = `<|assistant_end|>` (32763)

## Usage

Requires the `nanochat2` model type, available since [mlx-lm PR #1761](https://github.com/ml-explore/mlx-lm/pull/1761):

```bash
pip install mlx-lm
mlx_lm.generate --model lattice-research/lattice-quark-1.5b-mlx --prompt "what are you"
```

The chat template wraps prompts in the SFT marker format from training.

## Source

Base checkpoint: [lattice-research/lattice-quark-1.5b](https://huggingface.co/lattice-research/lattice-quark-1.5b)