Support matrix

This matrix follows the adapters and capability allowlists in the current code. “Supported” means integrated; it does not certify every model, device and platform combination through real-device acceptance.

Voice pipelines

Capability	Integrated	Cascaded
Processing	One realtime model handles speech directly	Speech recognition → language model → speech synthesis
Default combination	Qwen qwen-audio-3.0-realtime-plus	Volcengine ASR → DeepSeek deepseek-flash → Volcengine TTS
Speech interruption	Integrated	Integrated
Coding, search and memory tools	Available according to enabled capabilities	Available according to enabled capabilities
Camera frames in conversation	Not supported	Listed vision models only; off by default
Independent Loop Camera monitoring	Supported with a separate vision model	Supported with a separate vision model
iPhone text input	No editable input	Available when the host supports editable input

Integrated mode needs fewer settings. Use cascaded mode to change the language model or enable conversation vision. Both require credentials for the selected services.

Cascaded language models

Provider	Default model	Conversation vision	Credential
DeepSeek	deepseek-flash	Not exposed by the current Nova adapter	DEEPSEEK_API_KEY
Qwen / DashScope	qwen-plus	Not on the default model; select a listed Qwen vision model	DASHSCOPE_API_KEY
Volcengine Ark	doubao-seed-2-0-pro-260215	Supported	ARK_API_KEY

A model-name override does not grant image capability. A model supporting images is separate from Nova having a verified image-input adapter for it. Tool use also requires compatibility with the adapter's structured tool protocol.

Vision models

Provider	Models recognized by Nova
Qwen / DashScope	qwen3-vl-plus, qwen3-vl-flash, qwen-vl-max, qwen-vl-plus
Volcengine Ark	doubao-seed-2-0-pro-260215

Conversation vision uses the selected cascaded language model. Loop Camera uses a separate monitor-model connection, so an audio-only or DeepSeek foreground conversation can still delegate monitoring to a vision model. Unknown models do not silently fall back to another model.

Speech recognition and synthesis

Stage	Current adapter	Default resource / voice
ASR	Volcengine Speech	volc.seedasr.sauc.duration
TTS	Volcengine Speech	seed-tts-2.0 / zh_female_vv_uranus_bigtts
Integrated voice	Qwen realtime	longanqian

ASR uses DOUBAO_ASR_API_KEY when present, otherwise DOUBAO_BIGMODEL_API_KEY. TTS uses the latter. Supporting models and personal memory may still require DashScope credentials; see configuration.

Implementation: defaults, cascaded adapter selection, vision allowlist.