Qwen-AgentWorld-35B-A3B vs Qwen3-Omni-30B-A3B-Instruct

Qwen-AgentWorld-35B-A3B

A specialist, not a chat model. It plays the environment an agent acts in, across seven domains, so you can train or test an agent without the real thing. Alibaba ships BF16 only, which fits a 128 GB machine, and hasn't answered a discussion on it or changed the files since June.

Qwen3-Omni-30B-A3B-Instruct

An omni-modal cousin that listens, watches and reads, then answers in text or one of three voices, with tool calls from spoken requests, under plain Apache 2.0. Its only build, BF16, needs about 85 GB, and speech out needs Transformers or vLLM-Omni. The main risk is no answers from Qwen and no safety work for speech-driven actions.

Qwen-AgentWorld-35B-A3B

1. Use it as the environment, not the agent. Give it the domain's system prompt from Alibaba's GitHub repository, the history and the agent's action, and it returns the predicted observation
2. With vLLM pass --language-model-only, or startup fails on the missing vision weights
3. Use temperature 0.6, top_p 0.95 and top_k 20, and allow up to 32,768 output tokens, as the card recommends
4. Keep at least 128K of context, since the simulation depends on the history
5. Thinking is on by default and comes back through the qwen3 reasoning parser
6. A simulated observation is a prediction. Check anything that matters against the real environment

Qwen3-Omni-30B-A3B-Instruct

1. For text answers, mkdir -p media && vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --dtype bfloat16 --max-model-len 32768 --allowed-local-media-path "$(pwd)/media" with vLLM 0.11.1 or later serves the Thinker on an OpenAI-compatible API
2. The card's command passes --allowed-local-media-path /, which lets anyone who can reach the server have it read any file on the machine through a file:// URL. Point it at a dedicated folder that holds only the media you mean to send, as above, and give vLLM an absolute path
3. Send audio as audio_url and images as image_url content parts; vLLM serve doesn't take use_audio_in_video, so pass a video's audio separately
4. For speech out, load Qwen3OmniMoeForConditionalGeneration in Transformers 4.57.0 or later with qwen-omni-utils, pick speaker Ethan, Chelsie or Aiden, and write the returned waveform at 24 kHz
5. Call model.disable_talker() or pass return_audio=False when you only need text, which saves about 10 GB and time
6. Parse tool calls as JSON inside <tool_call> tags yourself, since no parser is named, and confirm actions triggered by speech before running them
7. Keep use_audio_in_video the same across every step of a conversation, the card warns results go wrong otherwise
8. Budget memory for video, the card's table rises from 78.85 GB for 15 seconds to 144.81 GB for 120 seconds in BF16