--- 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*