# LFM2.5-350M (cellm)

Liquid Foundation Model 2.5 (350M) converted to cellm format. Pre-quantized to 4-bit using MLX.

## Model details

- **Base model**: liquid-ai/LFM2.5-350M-MLX-4bit
- **Format**: cellm (.cellm)
- **Precision**: 4-bit weights (MLX int4), f16 norms
- **Size**: 211 MB
- **Architecture**: 16 layers, hybrid LIV convolution + grouped-query attention
  - Pattern: conv, conv, attn, conv, conv, attn, conv, conv, attn, conv, attn, conv, attn, conv, attn, conv
- **Hidden size**: 1024
- **Vocab**: 65,536

## Usage

```sh
./target/release/infer \
  --model lfm2.5-350m-v1.cellm \
  --tokenizer tokenizer.json \
  --prompt "The capital of France is" \
  --gen 40 --temperature 0 --backend cpu --kv-encoding f16
```

## Notes

- Text completion only. This is a base model.
- 350M parameters at 4-bit gives 211 MB on disk.
- The LIV (Linear Input-Varying) conv blocks use a causal sliding-window state per layer.
- Use temperature > 0 for more varied output.

## License

Apache 2.0 (same as upstream LFM2.5)