Tongyi DeepResearch 30B A3B

Tongyi DeepResearch 30B A3B is an advanced agentic large language model developed by Tongyi Lab. With 30 billion total parameters, it intelligently activates only 3 billion per token, making it highly efficient. This model is specifically optimized for long-horizon, deep information-seeking tasks and excels at complex agentic search, reasoning, and multi-step problem-solving, outperforming prior models on benchmarks like Humanity's Last Exam, BrowserComp, and GAIA. The model incorporates a fully automated synthetic data pipeline for scalable pre-training, fine-tuning, and reinforcement learning. It features large-scale continual pre-training on diverse agentic data to enhance reasoning and maintain currency. End-to-end on-policy RL with a customized Group Relative Policy Optimization ensures stable training. It supports ReAct for core ability checks and an IterResearch-based 'Heavy' mode for maximum performance. Ideal for advanced research agents and tool use, it offers a 131K token context window and 4K token max output. Pricing is competitive at $0.09/0.40 per 1M tokens (input/output).