Ai2 open source 8B scientific research report model: cutting out intermediate steps and speeding up generation by 3.5 times

Dongcha Beating AI News, Ai2 open source scientific research report model AstaBrief 8B. It is based on Qwen3-8B, which is the model responsible for "writing reports" in the Ai2 scientific research assistant Asta. Asta will first retrieve relevant papers and content fragments, and then hand them over to AstaBrief, which will organize these materials into a scientific research report with citations.

Asta's original Thinking model mainly relied on Claude and adopted a multi-step process: first summarizing and organizing the search results, and then generating reports into chapters. AstaBrief eliminates these intermediate steps. After getting the research questions and relevant paper fragments, you can directly generate a complete report in one go. Ai2 has now made it available for Asta's Fast mode.

Ai2 measured that it takes an average of 51.1 seconds to generate a report in Fast mode, while the original Claude-driven Thinking mode takes 178.5 seconds, which is about 3.5 times faster. In the ScholarQA-CS2 test, AstaBrief achieved citation precision of 90.5 and citation recall of 78.2; the original Qwen3-8B was 76.2 and 64.6, respectively.

The model weights are open using the Apache 2.0 protocol, and the training data is also made public. Developers can connect AstaBrief to their own paper retrieval system, or directly use official examples to process local PDFs. Ai2 also reminded that this batch of main training and evaluation was completed in 2025. The comparison was with the cutting-edge system at that time, and there was no complete re-test with the latest Claude and other models in 2026.