Qwen3-ASR 0.6B¶
Current status and review date¶
Decision snapshot
- Product role: Benchmark-qualified English one-shot STT challenger.
- Selected or best-qualified configuration: Official Qwen base image with vLLM 0.14.0 and the fail-closed official output parser, 16,384-token maximum model length, four maximum sequences, and normalized WAV input.
- Measured hardware: One RTX 5090 on Primary Node through an isolated endpoint; Qwen2.5-Omni remained co-resident and the RTX PRO 6000 was protected.
- Evidence: 3.621% primary-human micro-WER, 113.58 ms sequential p95, 137.36 ms concurrency-four p95, and zero final failures or repetitions.
- Decision:
challenger,no-promotion; it met the declared non-inferiority margin, but the retained record kept Parakeet routed. - Important limitation: Only English, non-streaming, one-shot transcription was qualified; multilingual behavior, streaming, translation, diarization, and VAD were not tested.
- Review dates: Retained evidence cutoff: 2026-07-28. Dossier-format review: 2026-08-31.
Review narrative¶
2026-07-28 — Runtime adaptation and smoke qualification¶
The first launch failed because the pinned runtime parsed
CUDA_VISIBLE_DEVICES as an integer, and the next launch rejected the custom
qwen3_asr architecture before the official package registered it. The
managed image retained Docker's device reservation, exposed ordinal 0
inside the restricted container, imported qwen_asr before vLLM, and applied
the official output parser fail-closed. The final endpoint returned valid
one-shot transcripts rather than Qwen's internal provider envelope.
2026-07-28 — Multi-sample corpus qualification¶
Qwen3-ASR completed the normalized English corpus sequentially and at concurrency four. Its 3.621% primary-human micro-WER met the declared non-inferiority margin, with lower p95 latency than the retained Parakeet baseline in this campaign. The evidence qualified a challenger, not a route change.
Immutable identity¶
Model¶
- Checkpoint:
Qwen/Qwen3-ASR-0.6B. - Revision:
5eb144179a02acc5e5ba31e748d22b0cf3e303b0.
Runtime¶
- Official base image digest:
sha256:fb75b775f089e06e5a1aaebffd421e37505cc630d50c86d889d95ffa45a7e16a. - Derived image tag:
anvil-qwen3-asr:official-vllm-0.14.0. - Derived image ID:
sha256:8bc0588b53f601ed2a8b9da23a89fa3d90f27bef70ea73d2efde2ba8f93ccc1c. - Engine: vLLM 0.14.0 with the official Qwen output parser.
Tested hardware and topology¶
Isolated candidate lane¶
- Host label: Primary Node.
- Measured device: one RTX 5090.
- Candidate endpoint: isolated, loopback-only serve.
- Co-resident workload: Qwen2.5-Omni remained running.
- Protected device: the RTX PRO 6000 was not used by the candidate.
Other accelerator products and multi-GPU execution were not tested.
Engine, quantization, KV, context, and concurrency recipe¶
Qualified recipe¶
- vLLM 0.14.0 with
qwen_asrimported before server startup. - Official
qwen_asr.parse_asr_outputapplied before returning the OpenAI transcription response. - Maximum model length: 16,384 tokens.
- Maximum sequences: four.
- Corpus schedules: three sequential repetitions and one concurrency-four run.
- Weight quantization and KV dtype: Not recorded in the retained dossier evidence.
The public experiment overlay and Compose service pin the reconstructable candidate configuration.
Evidence by measurement class¶
Functional¶
- One-shot transcription contract: pass after the fail-closed parser patch.
- Final failures and repetitions: zero.
Quality¶
- Primary-human micro-WER: 3.621%.
- Declared non-inferiority margin versus the retained baseline: pass.
Capacity and latency¶
- Sequential p50/p95: 67.40/113.58 ms.
- Concurrency-four p50/p95: 83.98/137.36 ms.
- Concurrency-four throughput: 47.25 requests/s.
Evidence¶
The dated finding links corpus provenance, sequential and concurrency-four results, startup identity, and the friction log.
Decision and promotion state¶
Qualified challenger¶
- Benchmark-qualified for the recorded English, non-streaming, one-shot STT contract.
- Met the declared non-inferiority margin.
Not authorized¶
no-promotion.- The retained decision kept Parakeet routed and does not imply current live state after the evidence cutoff.
Failures and gotchas¶
Resolved runtime failures¶
- The first launch failed on GPU-UUID parsing; the managed container retained
the Docker device reservation and exposed ordinal
0internally. - The second launch rejected the custom architecture; importing
qwen_asrbefore vLLM registered it. - The default 65,536-token context did not fit beside the protected workload; the qualified one-shot recipe pins 16,384.
- Raw provider envelopes were rejected until the official parser produced the transcript contract.
Unqualified behaviors¶
Multilingual recognition, streaming latency, translation, diarization, and VAD were not tested.