Qwen: Qwen3-ASR 1.7B
qwen/qwen3-asr-1.7bHugging FaceTranscriptionautomatic-speech-recognitioneval-resultsqwen3_asr
Providers
Live endpoints for this model. Response is the time from request to result; uptime is measured over the last day and week. Nothing here streams, so there is no first token to time and no token rate to report.
| Provider | Per audio hour | Response | Capacity | Uptime 1d | Uptime 7d | Health | Data |
|---|---|---|---|---|---|---|---|
| ImpCCbf1690% offCA | $0.0072 | 169ms30m | 8/8 free | 100.0% | 96.5% | Healthy | ZDR |
Pricing
USD per hour of audio input.
| Provider | Per audio hour |
|---|---|
| ImpCCbf1690% off | $0.0072 |
Price history
Price · USD per hour of audio
Performance
Percentiles cover the labelled window: the last 30 minutes when it has enough samples, otherwise the last day. Response is the time in seconds from request to result. This model answers whole rather than streaming, so there is no first token to time and no rate between tokens to report — response time is the whole measurement there is.
| Provider | Response p50 | p75 | p90 | p99 |
|---|---|---|---|---|
| ImpCCbf16 | 169ms30m | 208ms | 228ms | 800ms |
Rotating benchmark probes
Diagnostic scores from rotating dataset subsets. Percentages cover graded replies only and are not full benchmark results. Each score belongs to the provider’s serving configuration.
| Provider | fleurs | librispeech |
|---|---|---|
| ImpCCbf16 | 67.7%371 graded | 84.9%609 graded |
Benchmark answers are graded mechanically against the vendored dataset reference answers — numeric comparison of the final number for GSM8K, MATH-500 and AIME 2025, the committed multiple-choice letter for MMLU, MMLU-Pro, MMMU and MMBench, cosine ranking of the returned vectors for the Banking77 and SciDocs retrieval triplets, and per-utterance word match for LibriSpeech and FLEURS — never by another model; a reply that ran into the probe’s token cap before it committed to an answer is recorded but not graded.
Uptime
Share of successful requests and probes over the last day and week.
- ImpCCbf161d 100.0%7d 96.5%Healthy
Availability by hour · last 7 days
Share of successful requests and probes per hour and provider. The line spans hours with nothing to measure; where samples are sparse, a dot marks each hour actually measured.
Activity
Audio decoded per UTC day over the last 30 days. Every billed request counts, the platform's own paid probe traffic included.
curl https://api-staging.heimdallm.com/api/v1/audio/transcriptions \ -H "Authorization: Bearer $HEIMDALL_API_KEY" \ -F file=@interview.mp3 \ -F model="qwen/qwen3-asr-1.7b"Set HEIMDALL_API_KEY to a key from your account. See the quickstart for routing suffixes and provider preferences.
About
The Qwen3-ASR family includes Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, which support language identification and ASR for 52 languages and dialects. Both leverage large-scale speech training data and the strong audio understanding capability of their foundation model, Qwen3-Omni. Experiments show that the 1.7B version achieves state-of-the-art performance among open-source ASR models and is competitive with the strongest proprietary commercial APIs. Here are the main features:
- Author
- Qwen
- Family
- qwen/qwen3-asr-1.7b
- Weights
- Qwen/Qwen3-ASR-1.7B
- Parameters
- frequency_penalty (-2–2), max_completion_tokens (1–448), min_p (0–1), presence_penalty (-2–2), repetition_penalty (>0), response_format (json), seed (-9007199254740991–9007199254740991), temperature (0–1), top_k (≥-1), top_p (>0, ≤1)
- Licence
- apache-2.0
- Size
- 2.3B parameters, BF16 weights
- Tokenizer
- —
- Available until
- —
From the model card
What the repository states about these weights, read from Qwen/Qwen3-ASR-1.7B on Oct 11, 2026.
- Languages
- —
- Papers
- 2601.21337
- Datasets
- —
- Revision
- 7278e1eJan 30, 2026
- Downloads
- 1.1M
- Likes
- 1.2K