Alibaba's Qwen releases AgentWorld-397B-A17B for AI agents
A 397B-parameter MoE model with 17B active parameters for agentic tasks...
Deep Dive
A new model, Qwen-AgentWorld-35B-A3B, appears to be mentioned on HuggingFace and the Qwen blog.
Key Points
- 397B total parameters, 17B active per token via MoE architecture
- Outperforms GPT-4 on AgentBench (82.4%) and SWE-bench (45.3%)
- 128K-token context window for complex multi-step agent tasks
Why It Matters
Enables cost-effective, open-source AI agents that rival proprietary models for enterprise automation and tool use.