Speech-to-speech fine-tune of Qwen3-Omni-30B-A3B, quantized to 3.00 average bits per weight with DynQuant.

Fine-tuned from Qwen/Qwen3-Omni-30B-A3B-Instruct on spoken dialogue, then quantized to a 3.00-bit target with DynQuant.

Qwen3-Omni ships the codec decoder (code2wav) but no codec encoder or quantizer, so target speech cannot be turned back into the discrete Talker labels a speech-out loss would need. This run therefore trains the Thinker -- the half that listens and decides what to say -- inside the full speech-in/speech-out pipeline, with the Talker and code2wav frozen and bit-identical to the base checkpoint. That is also the only half DynQuant quantizes here: all 96 expert banks are Thinker banks.

There is no accuracy metric. No benchmark, no A/B against the base checkpoint, no listening study -- that was accepted going in. What did run is a smoke probe: 4 held-out clips per arm, seed pinned, each answered end to end from audio in to audio out, with the loaded module tree counted afterwards so that a directory which silently loaded unquantized could not pass as a working quantized model.

That checks for the destroyed-model failure -- silence, fluent nonsense, a load path that quietly fell back -- and all three arms pass it. It is not a measure of quality. A handful of replies read by eye cannot separate "as good as bf16" from "noticeably worse but still coherent", and nothing in this campaign does.

A 3.00-bit target is below this architecture's floor budget. The role floors alone cost 3.4210 average bits, so a 3.00 budget cannot pay for them and the soft floors bind: 394 of 650 modules are breached and 125 are cut to 2 bits.

The breaches land where they hurt a speech model most. The narrowest modules in the entire allocation are the audio tower's own attention projections -- audio_tower.layers.0.self_attn.{q,k,v}_proj at 2 bits against a 4-bit floor -- which is the path the model hears through. On a sibling arm of this same architecture a 3-bit budget cost 61.80 accuracy points and produced a destroyed model rather than a degraded one.

This artifact is published because it was asked for and because its size claim is real. The expectation going in was that it would be destroyed. The probe does not show that. Every reply came back fluent, on topic and grammatical, including one that correctly answered a spoken question about Plato's Republic -- which it could only do by hearing the question through those 2-bit projections.

It does show one behavioural difference. On the clip where the reference answer and both other arms decline a request for a person's home address, this arm offered to help find it. One clip is not a rate. But it is the kind of divergence that a fluency check is blind to by construction: instruction-following and refusal behaviour degrade before fluency does, so "it still speaks well" is the last thing to break and the weakest possible evidence that nothing else has.

So: not destroyed, not measured, and the weakest of the three. Listen before relying on it, and do not deploy it anywhere a refusal matters.

Produced with DynQuant v0.5.0. The allocation is the signal map the fine-tune itself recorded -- per-module gradient variance and activation RMS -- consumed by a role-aware ROI allocator. It was computed over the Thinker and applied to the full model, which is why the other two stages are still bf16.