Model reference · open weights
qwen3-asr is an open-weight audio or speech model from cstr, 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
Qwen3-ASR-1.7B — GGUF GGUF quantizations of Qwen/Qwen3-ASR-1.7B for use with CrispASR. Qwen3-ASR-1.7B is a multilingual speech recognition model supporting 30 languages + 22 Chinese dialects, with support for speech, singing voice, and songs with background music. Audio inputs up to 5 minutes. Files 2026-07 rebake — audio tower now Q80 (fixes empty / repeated transcripts) qwen3-asr-1.7b-q4k.gguf was re-baked with the 24-layer audio encoder kept at Q80 (previously Q4K like the LLM body; ~0.15 GB larger). The earlier Q4K export could emit an empty transcript (or, on longer audio, a repeated phrase) while still reporting success (CrispASR #240) — the same sub-8-bit encoder drift diagnosed in #218. Encoder-output cosine vs the F16 reference (jfk): 0.9632 → 0.9989 min, 0.9913 → 0.9998 mean — i.e. the tower is back to near-lossless and decode no longer degenerates. The Q80 and F16 files were never affected. Usage with CrispASR Performance On JFK speech sample (11s, English): - Q4K: 0.2x realtime on CPU (4 threads) - Perfect transcription accuracy Conversion Converted from the non-HF variant (Qwen/Qwen3-ASR-1.7B, not the -hf variant): Note: The converter also supports the -hf variant (Qwen/Qwen3-ASR-1.7B-hf) which uses a different tensor naming convention. Both produce identical results. Architecture - Audio encoder: 24-layer Whisper-style encoder (d=1024, 16 heads, GELU) - Projector: 2-layer MLP (1024 → GELU → 2048) - LLM decoder: 28-layer Qwen3 (d=2048, 16Q/8KV heads, headdim=128, SwiGLU) - Vocab: 151,936 tokens (GPT-2 BPE) - RoPE theta: 1,000,000 License Apache 2.0 Provenance and EU AI Act Art. 53 note - Upstream model: Qwen/Qwen3-ASR-1.7B — published by Qwen. - Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not. - What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs. - Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training da
Summarised from the published model card. Read the full card on the HuggingFace links below.
Specifications
| Maker | cstr |
|---|---|
| Type | Audio & music |
| Variants | 1 |
| Based on | Qwen/Qwen3-ASR-1.7B |
| Released | 2026-04-11 |
| Popularity | 8k downloads / month |
| Likes | 13 |
| 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.
| Variant | Params | Precision | VRAM | Fits 16 GB | Weights |
|---|---|---|---|---|---|
| qwen3-asr-1.7b-GGUF | — | GGUF | — | — | Weights ↗ |
Using it via the API
Once AxForge deploys cstr-qwen3-asr for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (cstr-qwen3-asr 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="cstr-qwen3-asr" -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 ↗