LocalClaw/TTS/Ossicle-4B
Local audio-language model

Ossicle-4B

Catalogue summary: English audio-language model for questions about speech and sound. Combines the Qwen3-Omni audio encoder with a Qwen3.5-4B backbone and merged LoRA weights. The complete checkpoint runs locally with PyTorch and Transformers; the single-pass ask mode returns text, not generated audio.

Repository editorial metadata; verify comparative claims in the linked upstream material.

GPU recommendedaudio-language question answering1 languagesApache 2.0
Choose an app

Start with the recommended app, or choose another supported option.

Compare speech models
Open model filesModel weights and official local setupOpen in UnslothLaunches Unsloth Desktop on this modelOpen Python packageVerified package page without exposing a commandOpen on Hugging FaceFiles, licence and model cardCheck my machineOptional: verify the fit with your saved hardware

Desktop app links require the app to be installed. If nothing opens, LocalClaw will show app-download and model-file fallbacks.

Catalogue quality
Not rated
Catalogue speed
Not rated
Model size
4.856B parameters; 9.724 GB safetensors
Voices
N/A: text answers; no speech synthesis, voice cloning or audio streaming

Can Ossicle-4B run locally?

Ossicle-4B can run locally for audio-language question answering over supplied recordings. Use the verified setup options on this page; no terminal command is required to choose the right path.

Apache 2.0 license. Still verify upstream usage notes before shipping.

Upstream source
audio-languagelong-form

Model details

Type
Local audio-language model
Family
ossicle
Latency
not rated
Formats
pytorchtransformerssafetensors
Languages
en
Context
Single-pass ask mode: author documents a 16 GB GPU configuration, 9.9 GB BF16 memory and 13.1 GB peak for 10 minutes of audio. The 16 GB comparison uses a 15 GiB PyTorch cap on an AMD RX 7900 XT 20 GB; it is not a measured system-RAM minimum. Transformers 5.17 tested upstream, trust_remote_code=True required. Chat and two-pass modes download Qwen3.5-4B separately; chat adds about 6 GB. Long-audio mode and 4-bit model loading remain in development. Only English is documented. Sources: HF model card and github.com/SILVERTHRONE32/ossicle-4b.

Install locally

01
Check runtimeConfirm the backend supports pytorch, transformers, safetensors on your machine.
02
Open recommended setupUse the app and model links above. LocalClaw does not expose a terminal command.
03
Test locallyRun a short private audio prompt before moving into production workflows.

Good for

  • audio-language question answering
  • GPU recommended local workflows
  • audio-language, long-form

Watch before shipping

  • Validate pronunciation, latency and artifacts with your own voice samples.
  • Review the upstream license and acceptable-use notes.
  • Benchmark on your target CPU, Apple Silicon or GPU setup.

Related TTS and speech models

OpenLLM France / LINAGORA Luciole 8B Audio 1.0 Local audio-language model · Q 9.1 · Speed 6.7 Kaiyang Li Mizar-159M Local audio-language model · Q 8.7 · Speed 8.8 jamiepine / Community Voicebox Local speech app · Q 9 · Speed 9.5 NVIDIA MagpieTTS Multilingual 357M Local TTS model · Q 9 · Speed 8.6 IndexTeam / Bilibili Index-Echo S2ST 2B Local speech translation model · Q 9 · Speed 7.8 Yang Xiao VeriSpoof-Plus Local audio-language model · Not rated Yang Xiao VeriSpoof-Base Local audio-language model · Not rated hexgrad Kokoro TTS Local TTS model · Q 9.2 · Speed 9.8
CompareBrowse all TTS models Local AIBrowse LLM models macOS appGet LocalClaw