Key takeaways

• The Tongyi DeepResearch 30B A3B is an agentic Large Language Model (LLM) designed for reasoning tasks, particularly quantitative and logical problem-solving.
• It utilizes a Mixture-of-Experts (MoE) architecture with 30 billion total parameters, activating 3 billion per token. Training involved a fully automated synthetic data pipeline, large-scale continual pre-training on diverse agentic data, and end-to-end on-policy Reinforcement Learning with Group Relative Policy Optimization.
• The model excels in arithmetic, algebraic manipulation, and logical deduction, while maintaining a low toxicity profile. However, it struggles with organic chemistry, biology/medicine, external context, nuanced physical constraints, and exhibits poor confidence calibration.
• The highest performance was on the AGIEval English dataset, achieving 84.82% accuracy, demonstrating proficiency in well-defined mathematical and logical problems.
• The worst performance was 9.17% accuracy on the Humanity's Last Exam (Biology/Medicine Knowledge Benchmark), indicating significant gaps in understanding nuanced biological and medical concepts and displaying overconfidence in incorrect answers.

Technical specifications

Context length

131k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~30B

Open weights

Yes

Licence

Open

Regions

China

Vendor

Qwen

Provider model key

alibaba/tongyi-deepresearch-30b-a3b