GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
QuantumDesk-AI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
QuantumDesk-AI
Model Tree
Qwen/Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
Native inference, greedy. v3 and v2 measured in the same run on the same audio: n=100 per WER set, and the 500-clip phone-quality half of our entity benchmark (synthetic caller utterances passed through 8 kHz resampling, telephone band-pass and μ-law/AMR compression).
| Metric | v3 | v2 |
|---|---|---|
| LibriSpeech test-clean WER ¹ | 2.17% | 2.12% |
| LibriSpeech test-other WER | 4.70% | 6.99% |
| Common Voice 17 Spanish WER | 4.17% | 30.2% |
| Phone-number digit accuracy | 97.0% | 95.5% |
| Person-name accuracy | 75.9% | 82.5% |
| Business-name accuracy | 83.8% | 86.8% |
| Spelled-email accuracy | 49.4% | 53.2% |
| Date / service accuracy | 100% / 99.6% | 100% / 100% |
¹ The synthetic training voices were cloned from LibriSpeech test-clean speakers (different sentences). Treat test-clean as slightly optimistic; test-other has no speaker overlap.
Trade-off: v3 is more robust on harder English audio, adds Spanish and is slightly better on phone numbers, but recognizes English person names about 7 points worse than v2 (business names and emails are within a few clips of v2). For English-only deployments where names matter most, v2 remains a good choice.
v3 responds natively to audio (not transcription-only), in the language the caller speaks. It is trained on this user turn shape:
text
[ {"role": "system", "content": "<your agent persona>"}, {"role": "user", "content": [ {"type": "text", "text": "Respond conversationally to what the user just said. Keep replies short — one or two natural sentences. Never use markdown, code, symbols, or headers. Never invent quotes. If you didn't catch what they said, ask them to repeat. "}, {"type": "input_audio", "input_audio": {"data": "<b64 wav>", "format": "wav"}} ]}]Transcription prompts:
"Repeat the following text, without any explanation: " + audio"Repeat the following text, which is written in Spanish, as it is without any explanation: " + audioA persona system prompt is strongly recommended for conversational use.
Three phases on 2×H200 (about $390 of compute in total):
The LLM is frozen throughout, so its text abilities are unchanged.
Works with vLLM (--trust-remote-code). Qwen 3.6's hybrid linear attention loses some
accuracy under vLLM relative to native transformers inference; for maximum quality use HF
transformers, or vLLM with --no-enable-chunked-prefill and default (fp32) mamba cache dtypes.
All numbers above are native.
Apache 2.0. Built with the Ultravox training framework by Fixie.ai.
[ {"role": "system", "content": "<your agent persona>"}, {"role": "user", "content": [ {"type": "text", "text": "Respond conversationally to what the user just said. Keep replies short — one or two natural sentences. Never use markdown, code, symbols, or headers. Never invent quotes. If you didn't catch what they said, ask them to repeat. "}, {"type": "input_audio", "input_audio": {"data": "<b64 wav>", "format": "wav"}} ]}]