Model reference · open weights
MOSS-SoundEffect is an open-weight audio or speech model from OpenMOSS-Team, listed in the AxForge catalogue. AxForge can bring it up on EU-owned hardware for you on request — with the licence handled where one is required.
About
MOSS-SoundEffect-V2.0 MOSS-SoundEffect v2.0 is a text-to-audio model with a Diffusion Transformer (DiT) backbone trained with the Flow Matching objective, paired with a DAC VAE and a Qwen3 text encoder. It generates high-fidelity environmental, urban, creature, and human-action sound effects from natural-language prompts, with controllable duration up to 30 seconds at 48 kHz. 1. Overview 1.1 TTS Family Positioning Within the MOSS-TTS Family, MOSS-SoundEffect is the dedicated text-to-sound model — the family member that turns natural-language captions into non-speech audio (ambience, urban scenes, creatures, human actions, short music-like clips). v2.0 supersedes the v1 discrete-token autoregressive backbone (MossTTSDelay) with a continuous-latent Diffusion Transformer + Flow Matching design. 1.2 Key Capabilities - Broad SFX coverage: natural environments, urban environments, animals & creatures, human actions, and short musical/percussive clips. - Long-form generation: stable audio up to 30 seconds per call with the duration tag prepended to the prompt at training time. - Bilingual prompts: trained with both English and Chinese captions. 1.3 Released Models Recommended inference hyperparameters 2. Quick Start Environment Setup We recommend a clean, isolated Python 3.12 environment to avoid dependency conflicts with the top-level MOSS-TTS environment. For a minimal inference-only install (still ships the Gradio demo; skips the fine-tuning extras): Basic Usage The underlying DiT is wrapped with torch.compile + Triton CUDA Graph. The first call may take a few minutes to compile. If you hit TorchDynamo / Triton errors, set TORCHDYNAMODISABLE=1 before launching Python. For a Gradio demo and fine-tuning recipes, see the GitHub README.
Summarised from the published model card. Read the full card on the HuggingFace links below.
Specifications
| Maker | OpenMOSS-Team |
|---|---|
| Type | Audio & music |
| Parameters (lead) | 1.4B |
| Variants | 2 |
| Runs with | diffusers |
| Based on | Qwen/Qwen3-1.7B |
| Released | 2026-05-25 |
| Popularity | 1k downloads / month |
| Likes | 136 |
| Licence | Open weights |
How it works
Variants
Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.
Using it via the API
Once AxForge deploys openmoss-team-moss-soundeffect for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (openmoss-team-moss-soundeffect below is illustrative; you get the exact model name on deployment.)
$ curl -sS https://api.axforge.ai/v1/audio/transcriptions \ -H "Authorization: Bearer $AXFORGE_API_KEY" \ -F model="openmoss-team-moss-soundeffect" -F file=@audio.mp3
Details
Languages
Tags
Licence
Open weights under apache-2.0 — commercial use is permitted. Deploy it on AxForge EU hardware on request. Read the licence ↗