Alibaba Releases New Open-Source AI Models
Alibaba has released two new open-source models, Tongyi Qianwen Qwen3.5-122B-A10B and Qwen3.5-35B-A3B, which reportedly excel in agent-based tasks and rival closed-source leaders. In other AI research, a new "Team of Thoughts" framework that orchestrates multiple diverse LLMs achieved 96.67% accuracy on difficult math tests at a lower cost than using a single massive model.
Alibaba's latest models utilize a "Mixture-of-Experts" (MoE) architecture, a key factor in their efficiency. This design only activates a fraction of the model's total parameters during any single task. For instance, the Qwen3.5-35B-A3B model has 35 billion parameters but only uses 3 billion for a given inference, allowing it to outperform its much larger 235-billion-parameter predecessor. The new models are designed for "agentic" workflows, where AI can plan and execute complex, multi-step tasks with minimal human input. To achieve this, they feature native tool-use capabilities for interacting with APIs and databases, and a massive 1-million-token context window, which allows for analysis of large documents or entire codebases without needing complex workarounds. This release is part of a broader strategy that has positioned Alibaba as a major player in open-source AI. The company's "Qwen" family of models has seen rapid adoption, with over 90,000 derivative models developed by the global open-source community on platforms like Hugging Face as of February 2025. This effort is a core component of the company's "user-first, AI-driven" strategic shift. The "Team of Thoughts" framework mentioned alongside Alibaba's release represents a different but complementary trend in AI research. Instead of relying on a single, massive model, this approach orchestrates multiple smaller, diverse AI models. This method is part of a larger family of reasoning techniques like "Chain of Thought" and "Tree of Thoughts," which guide models to break down complex problems into smaller steps to improve accuracy.
