What
Six architectures behind official recipes have no row in
.agents/model-matrix.md. (An earlier revision of this issue said seven. That was a miscount: the six
architectures below are reached by ten recipes, because MossTTSDelayModel
serves four of them. Six architectures, ten recipes — with #609's two, eight rows
in total.)
Unlike #609 these are not pin-lag — they are absent
from vLLM's core registry on main as well, because they live in the separate
vllm-project/vllm-omni repository.
We already track exactly one architecture on that basis:
MiniMaxH3DiTModel, whose row records that it "is a DIFFUSION architecture living
in the separate vllm-omni repository". The precedent exists; these seven simply
have no rows.
Confirmed present in vllm-omni/docs/models/supported_models.md:
| Architecture |
Recipes |
MossTTSDelayModel |
MOSS-TTS, MOSS-TTSD-v1.0, MOSS-SoundEffect, MOSS-VoiceGenerator |
MossTTSRealtime |
MOSS-TTS-Realtime |
Qwen3TTSForConditionalGeneration |
Qwen3-TTS-12Hz-1.7B-CustomVoice |
HiggsMultimodalQwen3ForConditionalGeneration |
Higgs Audio V3 TTS |
Not found in core main or the omni supported-models doc — these need placing
before they can be scoped, and the row should record where they actually live:
| Architecture |
Recipe |
VoxtralRealtimeForConditionalGeneration |
Voxtral-Mini-4B-Realtime-2602 |
BailingMMNativeForConditionalGeneration |
Ming-omni-tts-0.5B |
Why this is worth placing now
These are TTS / speech-generation models — an output modality we do not serve at
all. We ship ASR (Voxtral audio→text, Parakeet) and video generation (H3), so the
audio-generation gap is a real hole in the omni story rather than a long-tail
curiosity. Rows make the hole visible; they do not commit us to porting it.
Scope
Add the rows with upstream anchors (naming vllm-omni as the source repo, as the
MiniMaxH3DiTModel row does), INVENTORIED state, and update the CI-enforced
rollup counts in the same change. For the two unlocated architectures, the row
records target-pending with what was searched, rather than inventing an anchor.
Records-only.
Found while auditing recipes.vllm.ai coverage.
What
Six architectures behind official recipes have no row in
.agents/model-matrix.md. (An earlier revision of this issue said seven. That was a miscount: the sixarchitectures below are reached by ten recipes, because
MossTTSDelayModelserves four of them. Six architectures, ten recipes — with #609's two, eight rows
in total.)
Unlike #609 these are not pin-lag — they are absent
from vLLM's core registry on
mainas well, because they live in the separatevllm-project/vllm-omnirepository.We already track exactly one architecture on that basis:
MiniMaxH3DiTModel, whose row records that it "is a DIFFUSION architecture livingin the separate
vllm-omnirepository". The precedent exists; these seven simplyhave no rows.
Confirmed present in
vllm-omni/docs/models/supported_models.md:MossTTSDelayModelMossTTSRealtimeQwen3TTSForConditionalGenerationHiggsMultimodalQwen3ForConditionalGenerationNot found in core
mainor the omni supported-models doc — these need placingbefore they can be scoped, and the row should record where they actually live:
VoxtralRealtimeForConditionalGenerationBailingMMNativeForConditionalGenerationWhy this is worth placing now
These are TTS / speech-generation models — an output modality we do not serve at
all. We ship ASR (Voxtral audio→text, Parakeet) and video generation (H3), so the
audio-generation gap is a real hole in the omni story rather than a long-tail
curiosity. Rows make the hole visible; they do not commit us to porting it.
Scope
Add the rows with upstream anchors (naming
vllm-omnias the source repo, as theMiniMaxH3DiTModelrow does),INVENTORIEDstate, and update the CI-enforcedrollup counts in the same change. For the two unlocated architectures, the row
records
target-pendingwith what was searched, rather than inventing an anchor.Records-only.
Found while auditing
recipes.vllm.aicoverage.