Awesome model
first impression:
much better quality and accuracy than Qwen3.6-27B when handling long-term agent tasks (not 'coding' tasks).
Huge Thanks to Qwen Team !
Local Agent Model No.1
Can confirm it can indeed code and analyze code.
Edit:
However, it's still at 35B scores and also behind 27B
What tasks are you trying it on? On my coding tasks it performs better than Qwen3.6-35B but worse than 3.6 27B by a margin still. It should be noted that, according to the model card and paper, running it 'as an agent' really isn't what this model was trained for, but it is good to see that it can also be an agent and not just a world model.
I'm still wondering how to evaluate its world modeling capabilities on personal tests. Maybe side by side it with docker containers or something, at least for the terminal prompt.
What tasks are you trying it on? On my coding tasks it performs better than Qwen3.6-35B but worse than 3.6 27B by a margin still. It should be noted that, according to the model card and paper, running it 'as an agent' really isn't what this model was trained for, but it is good to see that it can also be an agent and not just a world model.
I'm still wondering how to evaluate its world modeling capabilities on personal tests. Maybe side by side it with docker containers or something, at least for the terminal prompt.
Yes, coding capabilities may require a much larger model, perhaps the 397B model instead. But for general agentic tasks, this 35B model performs better than Qwen3.6-27B in both quality and speed based on my personal tests.
Qwen 3.5 is not the right call to test, since Qwen 3.6 was a big update.
You're also comparing dense models to MOE ones, MOE models are always faster.
Qwen 3.5 is not the right call to test, since Qwen 3.6 was a big update.
You're also comparing dense models to MOE ones, MOE models are always faster.
Hey, you should test it by yourself, not arguing which model to compare. Just give it a try.
Tested on 2 x rtx pro 6000s,
Here are the docker-compose recipes if they help.
SGLang -- 200-240 TPS single request
services:
sglang:
image: lmsysorg/sglang:latest
container_name: sglang-qwen-agentworld
restart: unless-stopped
runtime: nvidia
ports:
- "18005:8000"
volumes:
- ~/.cache/huggingface:/root/.cache/huggingface
ipc: host
environment:
CUDA_DEVICE_ORDER: "PCI_BUS_ID"
NVIDIA_VISIBLE_DEVICES: "0,1"
NVIDIA_DRIVER_CAPABILITIES: "compute,utility"
SGLANG_ENABLE_SPEC_V2: "1"
command: >
sglang serve
--trust-remote-code
--model-path Qwen/Qwen-AgentWorld-35B-A3B
--served-model-name qwen-agentworld-35b
--tp 2
--mem-fraction-static 0.6
--context-length 262144
--reasoning-parser qwen3
--tool-call-parser qwen3_coder
--port 8000
--host 0.0.0.0
VLLM -- 80-100 TPS single request
services:
vllm:
image: vllm/vllm-openai:nightly
container_name: vllm-qwen-agentworld
ports:
- "18005:8000"
volumes:
- ${HOME}/.cache/huggingface:/root/.cache/huggingface
ipc: host
runtime: nvidia
environment:
- HUGGING_FACE_HUB_TOKEN=your-token
- CUDA_VISIBLE_DEVICES=0,1
- CUDA_DEVICE_ORDER=PCI_BUS_ID
command:
- --model
- Qwen/Qwen-AgentWorld-35B-A3B
- --served-model-name
- qwen-agentworld-35b
- --tensor-parallel-size
- "2"
- --gpu-memory-utilization
- "0.6"
- --max-model-len
- "262144"
- --max-num-seqs
- "8"
- --reasoning-parser
- qwen3
- --tool-call-parser
- qwen3_coder
- --enable-auto-tool-choice
- --language-model-only
Note that --language-model-only is needed as this model does not provide visual capabilities.
Guys, this is not an LM for personal use (chatbot), it's a world model one, it simulates a world state across several domains.
This is for example with the example system prompt given in the model card: You are a language world model simulating a Linux terminal environment. Given the user's command, predict the terminal output.
It will feel like a super slow computer lol but that can be a lot of fun to experiment with this! Possibilities are huge!
Guys, this is not an LM for personal use (chatbot), it's a world model one, it simulates a world state across several domains.
This is for example with the example system prompt given in the model card:
You are a language world model simulating a Linux terminal environment. Given the user's command, predict the terminal output.
It will feel like a super slow computer lol but that can be a lot of fun to experiment with this! Possibilities are huge!
Look at this agent-loop demo, driven by Qwen-Agentworld-35B-A3B, it's an agentic model and also can be used in general conversations.
Efficient.
Cool if that works but that's not how the model is supposed to be used... Nevertheless, interesting side effect that it works.
Is it worth to use it in comparison with 3.6 35 a3b. This is finetune of 3.5 version and 3.5 has significantly lower capabilities
Worth noting the developers at Agents A1 had a very great take on this. Model is much more refined with vision. Also indeed incredibly fast as Agentic toolcaller, just as bougee with code.