# Qwen3-0.6B (cellm)

Qwen3-0.6B base model converted to cellm format at various quantization levels.

This is a base (pretrained) model, not instruction-tuned. It works in text completion mode but does not follow chat templates.

## Model details

- **Base model**: Qwen/Qwen3-0.6B
- **Architecture**: Standard transformer with grouped-query attention
- **Hidden size**: 1024
- **Attention**: 16 heads, 128 head dim, attn_dim = 2048
- **Vocab**: 151,936

## Variants

| File | Format | Size | Quality |
|------|--------|------|---------|
| qwen3-0.6b-f16.cellm | f16 | 1.4 GB | Repetitive but on-topic |
| qwen3-0.6b-f16-new.cellm | f16 (rebuilt) | 1.4 GB | Same as above |
| qwen3-0.6b-int8.cellm | int8 | 866 MB | Repetitive but on-topic |
| qwen3-0.6b-int4-g32.cellm | int4 g32 | 830 MB | Noisy, sometimes on-topic |
| qwen3-0.6b-mixed-int4.cellm | mixed int4 | 810 MB | Mostly repetitive |
| qwen3-0.6b-int4-g128.cellm | int4 g128 | 590 MB | Mostly repetitive |
| qwen3-0.6b-test-i4.cellm | int4 test | 582 MB | Fragmented |
| qwen3-0.6b-int4.cellm | int4 aggressive | 360 MB | Fragmented |

All variants share one tokenizer.json.

## Usage

```sh
./target/release/infer \
  --model qwen3-0.6b-f16.cellm \
  --tokenizer tokenizer.json \
  --prompt "Explain what a computer is:" \
  --gen 40 --temperature 0.7 --backend cpu --kv-encoding f16
```

## Notes

- This is a 0.6B base model. Output quality is limited across all quantizations.
- Use temperature 0.7+ to avoid repetition loops.
- f16 and int8 are the most coherent. Below int8, quality drops.
- For instruction-following, use Qwen3.5-0.8B instead.

## License

Apache 2.0 (same as upstream Qwen3)