# Qwen3-ASR-0.6B (slim)

> Qwen3-ASR-0.6B is Alibaba's compact open-weight speech recognition model for 30 languages and 22 Chinese dialects, with language identification, context hints and streaming, published on Hugging Face in BF16 under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-asr-0-6b.md (~10,850 tokens) · this version ~2,380 tokens · JSON https://www.anchorterminal.com/tools/qwen3-asr-0-6b.json · canonical https://www.anchorterminal.com/tools/qwen3-asr-0-6b
- 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-asr-0-6b#pricing · reviews https://www.anchorterminal.com/tools/qwen3-asr-0-6b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61.2/100 · rank #32 of 100 open-weight models · #15 in Speech-to-text · not agent-ready · confidence medium**


Assessment: A very small local transcription cousin for 30 languages and 22 Chinese dialects, built for throughput and realtime streaming through vLLM, under plain Apache 2.0. Its BF16 files need about 2 GB and run in released vLLM, SGLang, llama.cpp and Transformers. It trades some accuracy for speed, and has no publisher replies or safety results.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Speech-to-text · legal entity: Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3-ASR-0.6B`, huggingface `Qwen/Qwen3-ASR-0.6B-hf`
- 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-ASR-0.6B and Qwen's Transformers-native copy Qwen/Qwen3-ASR-0.6B-hf, read from the model cards (the original card is shared with the 1.7B model), the files and their metadata, the chat templates, the QwenLM/Qwen3-ASR GitHub repository and technical report, and the runtimes' own code and documentation on 21 October 2026. Nothing was downloaded or run. The hosted Qwen3-ASR-Flash on Model Studio is a separately named model and isn't graded
- Model: An AuT audio encoder of about 180M parameters (18 layers, hidden size 896, 128 mel bins) with a projector into Qwen3-0.6B, per the technical report. The card names Qwen3-Omni as its foundation model. Audio in, text out. The -hf files count 782,426,112 parameters; the original files store 938,008,576 because they keep a copy of the tied output embedding. The card states no parameter figure
- Languages: Language identification and recognition for 30 languages and 22 Chinese dialects, for speech, singing and songs with background music, per the card
- Files: One safetensors file in BF16 (1.88 GB for the whole repository) with chat_template.json, and a 1.56 GB safetensors file with chat_template.jinja in the -hf repository. No pickle files and no Python files. Hugging Face lists a SHA-256 for every LFS file. No LICENSE file in either Hugging Face repository; the card metadata says apache-2.0 and the GitHub repository carries the Apache 2.0 text
- Licence: Apache 2.0, with no use policy in the repositories or on the card
- Runtimes: vLLM (Qwen3ASRForConditionalGeneration in v0.32.0, plus Qwen3ASRRealtimeGeneration for realtime streaming of this size), SGLang (qwen3_asr in v0.5.22), llama.cpp (converter and qwen3a audio projector in release b11772, run from ggml-org's GGUF builds) and Transformers (qwen3_asr since 5.13.0, loading the -hf repository). Qwen's own qwen-asr package wraps Transformers and vLLM and adds the forced aligner. Not in Ollama's or LM Studio's libraries
- Input and output: 16 kHz audio as a path, URL, base64 or array, with an optional forced language and free-text context in the system message. Output is "language <name><asr_text><transcript>", which parse_asr_output in qwen-asr or the processor's decode with return_format parsed splits into language and text. Word or character timestamps need Qwen3-ForcedAligner-0.6B. Streaming runs through the vLLM backend only, without batching or timestamps. The technical report gives single recordings up to 20 minutes
- Context: 65,536 tokens (max_position_embeddings in config.json). The card states no context length
- Recommended settings: BF16, FlashAttention 2 where available, greedy decoding, max_new_tokens 256 for short clips and more for long audio, per the card
- Hardware: The card recommends a GPU that runs FlashAttention 2 and shows CUDA in every example. It states no memory figure; the BF16 files are 1.6 to 1.9 GB
- Training data: The technical report describes about 40 million hours of pseudo-labelled ASR data for encoder pretraining (mostly Chinese and English), 3 trillion tokens of multimodal pretraining, a smaller disjoint multilingual SFT set, and about 50,000 utterances for reinforcement learning. No sources, cut-off or filtering are given

- Scores: Capability 59, Reach 75, Risk 55, Agent interface 75, Documentation and upkeep 33 · total over the 7 assessed categories
- Why: Capability, From 3 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns 30 because Alibaba's smallest build, the -hf BF16 repository, is 1.58 GB, about 1.9 GB with the fifth added, well within 16 GB. · Risk, File safety gets 25, safetensors only and loaded by the runtimes' own code with a Jinja template. · Agent interface, Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The shared card covers intended use, architecture and languages and misses limitations, a parameter figure, context length and a cut-off (8)…
- Sources: 20, open questions: 5, both in the full twin
- Capabilities: speech.stt, speech.batch, speech.streaming, speech.languages, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-asr-0-6b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-asr-0-6b.svg` or a link to https://www.anchorterminal.com/tools/qwen3-asr-0-6b 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. Start it with `vllm serve Qwen/Qwen3-ASR-0.6B` (vLLM v0.32.0 or later), then POST the WAV to /v1/audio/transcriptions or send an audio_url part to /v1/chat/completions
2. For vLLM's realtime streaming, start it with `--hf-overrides '{"architectures":["Qwen3ASRRealtimeGeneration"]}'`, which vLLM doesn't pick from config.json
3. Split chat-completions output with `parse_asr_output` from qwen-asr, since it returns `language <name><asr_text>` before the transcript
4. For Transformers, load Qwen/Qwen3-ASR-0.6B-hf with Transformers 5.13.0 or later and use `processor.apply_transcription_request(audio, language=..., prompt=...)`
5. Put names and domain terms in the system message or `prompt` to bias the transcript, and force the language when you know it
6. For timestamps, use `pip install -U qwen-asr` and pass `forced_aligner="Qwen/Qwen3-ForcedAligner-0.6B"` with `return_time_stamps=True`
7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions

## Connect

```bash
pip install -U "qwen-asr[vllm]"
qwen-asr-serve Qwen/Qwen3-ASR-0.6B --host 0.0.0.0 --port 8000
# or, with vLLM itself:
vllm serve Qwen/Qwen3-ASR-0.6B
```

```bash
curl http://localhost:8000/v1/audio/transcriptions \
  -F model=Qwen/Qwen3-ASR-0.6B \
  -F file=@/path/to/speech.wav
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.batch, speech.streaming, 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.batch, speech.streaming, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Nemotron 3.5 ASR Streaming 0.6B | C | 59.8 | speech.stt, speech.streaming, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.min.md |
| Amazon Transcribe | BB | 75.3 | speech.stt, speech.streaming, speech.batch, speech.languages | https://www.anchorterminal.com/tools/amazon-transcribe.min.md |
| Azure AI Speech speech-to-text | BB | 73 | speech.stt, speech.streaming, speech.batch, speech.languages | https://www.anchorterminal.com/tools/azure-speech-to-text.min.md |


## Panel reviews (4, average 3/5, desk reviews from public material, no calls made)

- ★★★☆☆ A dated note for the -hf build, and no replies (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 1.58 GB of BF16 for transcription (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Parsed transcripts, timestamps from a second model (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A Docker example that mounts the host's socket (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 4. They agree the safetensors load with runtime code under Apache 2.0, that Alibaba documents the inputs and parsed output, and that no Qwen member has answered its discussions. Read it as a 1.58 GB transcriber with context hints and four runtimes, whose card's Docker example mounts the host's socket without a caution.