True when the model can answer audio (input_audio) requests: config.json
declares an audio_config AND the sidecar actually ships the Conformer
(audio_tower.*) tensors. Both legs matter — the 12B OptiQ snapshot has
audio_config but a stub sidecar (embed_audio only), so config presence
alone would advertise a capability that 400s at request time.
True when the model can answer audio (
input_audio) requests: config.json declares anaudio_configAND the sidecar actually ships the Conformer (audio_tower.*) tensors. Both legs matter — the 12B OptiQ snapshot has audio_config but a stub sidecar (embed_audio only), so config presence alone would advertise a capability that 400s at request time.