Qwen3-tts

Qwen3-TTS raw-ELF model for LLiMa

This repository contains the model assets only for Qwen3-TTS on the SiMa Modalix DevKit. Download or copy the directory intact and pass its path to llima run.

Requirements

• ARM64 SiMa Modalix DevKit running platform 2.1.3.
• Matching LLiMa packages installed on the board: sima-lmm-core, sima-lmm-cli, and sima-lmm-qwen3tts-runtime.
• No CSM or Python/Conda environment is required on the board. The installed runner package supplies qwen3tts and its matched ARM64 Torch, NumPy, and OpenBLAS shared-library closure.

Contents

• qwen3_model/mpk/*.elf — raw compiled MLA model ELFs.
• qwen3_model/devkit/codec_tail_raw_mla_contract.json — codec-tail contract.
• qwen3_components/ — tokenizer and safetensor assets.
• devkit/qwen3_tts_config.json — model configuration.

Run

llima run /path/to/Qwen3-tts \
  --prompt "Hello from Qwen3-TTS on the devkit." \
  --speaker Vivian \
  --language English \
  --output-wav output.wav

Long utterances

The raw codec-transformer and codec-tail artifacts each use a fixed 50-frame transport ABI. The installed runner processes both stages in sequential 50-frame windows, enabling utterances beyond that boundary (up to the runner's --max-frames 512 safety limit). This behavior was validated on a Modalix DevKit with a 128-frame generation and WAV output.

The streaming endpoint is enabled by default. For deterministic maximum-frame validation, invoke /usr/lib/aarch64-linux-gnu/sima-lmm/qwen3tts/bin/qwen3tts directly with --endpoint-disable.