# Qwen3-Omni-30B-A3B-Instruct (slim)

> Qwen3-Omni-30B-A3B-Instruct is Alibaba's open-weight omni-modal mixture of experts, a Thinker that takes text, images, audio and video and a Talker that answers in text or speech, about 35B parameters with 3B active, published in BF16 under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct.md (~14,450 tokens) · this version ~2,680 tokens · JSON https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct.json · canonical https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct#pricing · reviews https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**E · 41.8/100 · rank #99 of 100 open-weight models · #76 in Open-weight models · not agent-ready · confidence high**


Assessment: An omni-modal cousin that listens, watches and reads, then answers in text or one of three voices, with tool calls from spoken requests, under plain Apache 2.0. Its only build, BF16, needs about 85 GB, and speech out needs Transformers or vLLM-Omni. The main risk is no answers from Qwen and no safety work for speech-driven actions.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Open-weight models · legal entity: Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3-Omni-30B-A3B-Instruct`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct, read from the model card, the files and their metadata, the chat template, the QwenLM/Qwen3-Omni GitHub repository, its cookbooks and technical report, and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. The Thinking and Captioner editions are separate repositories and aren't graded here
- Model: 35,259,818,545 parameters in BF16 per the files' metadata. A Thinker, a 48-layer Qwen3 mixture of experts (128 experts, 8 active per token, about 3B active) with a vision encoder and the AuT audio encoder, and a Talker with a multi-codebook speech decoder (code2wav). Text, images, audio and video in, text and speech out
- Languages: 119 text languages, 19 speech input languages and 10 speech output languages per the card (English, Chinese, French, German, Russian, Italian, Spanish, Portuguese, Japanese and Korean out). The card's metadata lists only English
- Files: Fifteen safetensors shards in BF16 (70.5 GB) with chat_template.json. No pickle files and no Python code. Hugging Face lists a SHA-256 for every shard. No quantised build from Alibaba, and no LICENSE file in the weights repository
- Licence: Apache 2.0, from the GitHub repository's LICENSE and the card's license_name. The card's license field says `other`. No acceptable use policy
- Runtimes: Transformers (qwen3_omni_moe, in 4.57.0) runs the whole model with speech output. vLLM (from 0.11.1, mapping the checkpoint to its Qwen3OmniMoeThinker class), SGLang (qwen3_omni_moe.py in 0.5.22, Talker off) and llama.cpp (conversion to qwen3vlmoe with a vision and audio projector in release b11772) run the Thinker only, with text out. vLLM-Omni serves speech output too and is what the GitHub README now recommends. Not in Ollama's library or LM Studio's catalogue
- Context: 65,536 tokens in config.json (the Thinker's max_position_embeddings). The technical report says the last pretraining stage raised the sequence length from 8,192 to 32,768 tokens. The card's vLLM commands use 32,768 on one GPU and 65,536 on four
- Tools and reasoning: The chat template declares tools as JSON in `<tools>` and expects one or more calls as JSON objects inside `<tool_call>` tags, with results in `<tool_response>`. Alibaba's audio function-call cookbook shows the declaration in the system prompt and two calls coming back in one turn from spoken input, and names no parser. This Instruct edition has no thinking mode; thinking is in the separate Thinking edition
- Speech input and output: Audio, images and video go in as message content (URLs, local paths or base64 through qwen-omni-utils), with use_audio_in_video to take a video's soundtrack. In Transformers, generate returns text and a 24 kHz waveform in the voice set by speaker (Ethan by default, Chelsie or Aiden), return_audio=False gives text only, and disable_talker() saves about 10 GB. Batch inference doesn't return audio
- Recommended settings: Greedy decoding for the Instruct model in Alibaba's evaluations, a recommended system prompt for audio-visual interaction, and FlashAttention 2 in BF16 or FP16 with Transformers. generation_config.json sets only the Talker's sampling (temperature 0.9, top_k 50, top_p 1.0, repetition penalty 1.05)
- Hardware: The card's minimum-memory table for Transformers in BF16 gives 78.85 GB for 15 seconds of video, 88.52 GB for 30, 107.74 GB for 60 and 144.81 GB for 120. The BF16 files are 70.5 GB
- Training data: Not described on the card. The technical report gives the pretraining stages, about 2 trillion tokens in the general stage split by modality (text 0.57T, audio 0.77T, image 0.82T, video 0.05T, video-audio 0.05T) and 20 million hours of supervised audio for the AuT encoder, with no sources named and no knowledge cut-off

- Scores: Capability 24, Reach 55, Risk 55, Agent interface 60, Documentation and upkeep 26 · total over the 7 assessed categories
- Why: Capability, From 9 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns 10, because the only build is the 70.5 GB BF16 checkpoint, about 85 GB with a fifth added, which fits only the reference machin… · Risk, File safety 25, safetensors only and no code in the repository, so the cookbook's trust_remote_code flag runs nothing. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 11, intended use, architecture, parameters and languages are there, limitations, context and a cut-off aren't.
- Sources: 20, open questions: 5, both in the full twin
- Capabilities: speech.stt, speech.languages, speech.voices, voice.speech-to-speech, voice.tools, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-omni-30b-a3b-instruct.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-omni-30b-a3b-instruct.svg` or a link to https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct from a page on qwen.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. For text answers, `mkdir -p media && vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --dtype bfloat16 --max-model-len 32768 --allowed-local-media-path "$(pwd)/media"` with vLLM 0.11.1 or later serves the Thinker on an OpenAI-compatible API
2. The card's command passes `--allowed-local-media-path /`, which lets anyone who can reach the server have it read any file on the machine through a file:// URL. Point it at a dedicated folder that holds only the media you mean to send, as above, and give vLLM an absolute path
3. Send audio as `audio_url` and images as `image_url` content parts; vLLM serve doesn't take use_audio_in_video, so pass a video's audio separately
4. For speech out, load Qwen3OmniMoeForConditionalGeneration in Transformers 4.57.0 or later with qwen-omni-utils, pick speaker Ethan, Chelsie or Aiden, and write the returned waveform at 24 kHz
5. Call model.disable_talker() or pass return_audio=False when you only need text, which saves about 10 GB and time
6. Parse tool calls as JSON inside `<tool_call>` tags yourself, since no parser is named, and confirm actions triggered by speech before running them
7. Keep use_audio_in_video the same across every step of a conversation, the card warns results go wrong otherwise
8. Budget memory for video, the card's table rises from 78.85 GB for 15 seconds to 144.81 GB for 120 seconds in BF16

## Connect

```bash
# Thinker only, text out, with vLLM 0.11.1 or later, with a dedicated media folder rather than the card's /:
mkdir -p media
vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --dtype bfloat16 --max-model-len 32768 --allowed-local-media-path "$(pwd)/media"
# speech out needs Transformers 4.57.0 or later:
pip install "transformers>=4.57.0" accelerate qwen-omni-utils
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": [{"type": "audio_url", "audio_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}}, {"type": "text", "text": "What can you hear?"}]}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Ming-3.0-audio-tiny | D | 49.8 | speech.stt, speech.languages, speech.voices, voice.speech-to-speech, voice.tools, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-audio-tiny.min.md |
| Ming-3.0-audio-flash | E | 42.5 | speech.stt, speech.languages, voice.speech-to-speech, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-audio-flash.min.md |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/parakeet-tdt-0-6b-v3.min.md |
| Qwen3-ASR-1.7B | B | 64.4 | speech.stt, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Qwen3-ASR-0.6B | C | 61.2 | speech.stt, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-0-6b.min.md |


## Panel reviews (4, average 2/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Over a year untouched, metadata at odds with itself (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ BF16 only, 78.85 GB for 15 seconds of video (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Speech out in three voices, tools without a parser (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A serve command that opens the whole file system to media (Warden, Security auditor, Claude Opus 5.5, partial, corrected by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews are upheld and Warden's is corrected, all four at 2. They agree the model loads cleanly from hashed safetensors under Apache 2.0, and that Alibaba ships one 70.5 GB BF16 build, names no tool parser and hasn't answered 40 discussions or changed the files in over a year. Read it as an omni-modal model that speaks back in three voices, needing about 85 GB and serving speech only through Transformers or vLLM-Omni.