File size: 2,868 Bytes
fca9437
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
---
library_name: haru
pipeline_tag: audio-to-audio
base_model:
- higurashi-ai/Haru-4B
---

This is a randomly initialized tiny model based on [higurashi-ai/Haru-4B](https://huggingface.co/higurashi-ai/Haru-4B) for testing and debugging.

| Feature | Source | Tiny |
|---|---|---|
| Checkpoint tensor elements | 5.27B BF16 | 8.36M BF16 |
| Safetensors weight files | 10.54 GB (model.safetensors) | 16.74 MB (model.safetensors) |
| Language model | Qwen3.5-4B, 32 layers | Qwen3.5, 8 layers |
| Attention cadence | 3 linear + 1 full | Same; two complete cycles |
| Hidden / MLP width | 2,560 / 9,216 | 32 / 64 |
| Q heads / KV heads / head dimension | 16 / 4 / 256 | 4 / 1 / 16 |
| Vocabulary / tied embeddings | 248,320 / true | 248,320 / true |
| Codec | VibeVoice, base 32, depths 3-8 | VibeVoice, base 1, depth 1 |
| Audio latents | 64 dims at 7.5 Hz | 64 dims at 7.5 Hz |
| MeanFlow head layers / steps | 4 / 1 | 1 / 1 |

| File path | Size |
|------|------|
| model.safetensors | 16.7MB |

### Example usage

#### Haru duplex session (✅ verified)

Install with `pip install 'git+https://github.com/higurashi-ai/Haru#subdirectory=haru'`. Runs six 133 ms frames with a synthetic reference voice; only the language model's linear-attention layers need CUDA, the codec and head also run on CPU.

```python
def run_haru(model_id: str) -> None:
    import numpy as np
    import torch
    from haru import DuplexConfig, DuplexSession, HaruModel

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = HaruModel.from_pretrained(model_id, device=device)
    seconds = np.arange(24_000) / 24_000
    voice = (0.2 * np.sin(2 * np.pi * 150 * seconds)).astype(np.float32)
    config = DuplexConfig(seed=7, speak_first=True, max_frames=6)
    session = DuplexSession(model, "You are Haru. Greet the other person.", voice, config)
    frames = [session.start()]
    while not frames[-1].done and len(frames) < config.max_frames:
        frames.append(session.step(np.zeros(3200, dtype=np.float32)))
    assert len(frames) == 6
    assert all(frame.audio.shape == (3200,) and frame.audio.dtype == np.float32 for frame in frames)
    assert all(np.isfinite(frame.audio).all() and np.abs(frame.audio).max() <= 1.0 for frame in frames)
    assert all(0 <= frame.token < 248320 for frame in frames)
    assert frames[1].token not in (248076, 248071)
    print("frames", len(frames), "text", repr("".join(frame.text for frame in frames)[:80]))


if __name__ == "__main__":
    run_haru(model_id="tiny-random/haru")
```

### Creating this repo

- Code: [.meta/create_model.py](.meta/create_model.py)
- Model structure: [.meta/model.log](.meta/model.log)
- Environment: [.meta/env.json](.meta/env.json)

### Change log

- 2026-11-15: Tiny Haru full-duplex checkpoint with a shrunken Qwen3.5 backbone, codec and MeanFlow head.

*This repo is generated by GPT-6.1-Sol*