Feature Compatibility

Feature compatibility

Use this guide to choose engines, input types and CPU/GPU devices for your task. Each model family has its own supported modalities, runtime dependencies and memory requirements. Read it alongside the individual workspace pages. Installation and worker readiness describe different states from successful generation; the sections below explain the requirements for each task.

CPU and GPU support

The workspace, API and CLI do not require CUDA. CPU-capable models and workflows can use the CPU, subject to system RAM and the selected engine's requirements. CPU generation can be much slower than GPU generation.

ComfyUI provides a CPU execution mode. Omni's release checks started ComfyUI explicitly on CPU and discovered its nodes with an empty queue. That establishes CPU startup and integration; generation still depends on the chosen model and nodes. Some engines, quantization kernels and custom nodes specifically require CUDA. Check the requirements for the task you intend to run.

Python is bundled for both sides of the portable app: Linux Python and model environments inside the distro, and Windows Python for the networking relay. No separate Python installation is required.

Speech workspaces

The Voice & TTS nav item and Home card are labelled Soon. There is no engine picker and no Generate control. The MOSS page has speech controls; see the decoder compatibility notes below before using them. Do not wait on this tab during a session.

MOSS-TTS decoder compatibility

Earlier MOSS-TTS runs saved 48 kHz WAVs. On the tested build, a fresh worker loaded but inference failed in the upstream decoder with a CUDA driver error and WSL dxgresource exhaustion. Omni retires that worker after the failure. Do not keep restarting it as a capability probe. MOSS-SoundEffect remains a separate, qualified path. Previously saved speech or a recorded vocal can still be used as ACE-Step source audio.

Moshi has no usable batch or stream path

Moshi weights may be installable for a future duplex transport. The current API:

• batch infer raises 501
• Moshi is absent from the stream-handler registry
• Chat hides Moshi
• Testing hides Moshi from batch load

Loading VRAM cannot make Chat or /api/chat/moshi work. Do not spawn Moshi as a capability probe; it wastes 16 GB class VRAM and still 501s.

Qwen input and output types

Qwen2.5-Omni text is verified (3B; 7B GPTQ-int4 text/image on 24 GB). The worker does not wire audio or video. Those fields return 501 instead of ignoring the file. Native Talker/TTS is not registered on /api/tts. Official upstream speech does not make Omni's route work.

Chat may still show capability hints derived from model descriptions. Trust this page and the disabled-attachment behavior when it matches; if an attach button is enabled and the server 501s, the server is the source of truth.

Qwen3-Omni 30B: audio/video also 501, and the installed BF16-class variants hard-stop before spawn on a typical 24+12 GB host (valid: false).

MiniCPM-o: understanding yes, speech and video no

Verified on a 24 GB GPU: text, small-image description, 16 kHz audio understanding (one modality per request, audio ≤ 60 s).

Unavailable:

• Video — explicitly rejected
• Combined image+audio — 400
• Native TTS (/api/tts/minicpm_o) — decoder incompatibility (get_mask_sizes). Do not use it.

AnyGPT is not a live multimodal worker

The worker implements text only and ignores speech/music/image fields. Cold load has stayed CPU-side for many minutes, retained ~16 GB RAM, and missed the bridge window. Delete the worker if you started one by mistake. Do not autospawn AnyGPT to "see if it works now."

ACE specialized adapters

Music sub-tabs Lyric→Vocal and Text→Samples depend on official adapters that are unreleased. The registry refuses false installs. CLI ace-step lyric2vocal / text2samples will not become working because you passed --prompt.

If a future status row shows those adapters installed with enables_mode, revisit this paragraph. Until then: Generate with lyrics, Vocal→BGM, Audio Lab, and MOSS-SFX.

The legacy Chinese Rap LoRA pack is PEFT-incompatible and rejected even if files are on disk.

ACE extend/complete can rewrite audio you hoped to keep. Use composition for stable assembly.

Music 3 task scope

Official Diffusers subset only. No verified LoRA attach, no voice cloning, no streaming, no reference audio. Lyrics are required even for instrumentals. Load plans that are valid: false are refusals, not hints.

Oversized 30B Omni variants

Qwen3-Omni ~60 GB and Nemotron BF16 ~62 GB exceed a 24+12 GB safe budget. Analysis returns an invalid plan and spawn hard-stops before process creation. Nemotron NVFP4 (~21 GB) is the candidate that should prefer the 24 GB card alone; it is not live-verified here. Do not download 60 GB checkpoints expecting a miracle device map.

GPTQ Qwen uses single placement; automatic multi-GPU placement is unavailable.

Comfy: analysis is not generation

A green requirement check means files and nodes look present. A valid placement plan means the estimate fits current free VRAM. It does not mean the graph will finish at 4K, 121 frames, or temporal_size=2048. Staged LTX decode has OOMed a 12 GB card on aggressive temporal settings. Analyze-before-run, then start small.

prompt_id is not a PNG. valid: false is a hard stop.

Core update and Manager update_all are different buttons. Do not skip verification of each.

Jobs versus output ids

Polling /api/jobs for ACE/Audio Lab/Music 3 generate is wrong. Poll only when the response says the operation is running (installs, compose, maintenance).

Tokens and security

Never print X-Omni-Token, omni-cli session show, or a newly created bearer key in logs, tickets, or git. Loopback does not mean "safe to paste in Discord."

Related pages

• Voice & TTS
• Chat
• Music
• GPU placement