Open-sourcing AstaBrief, the fast report-generation model in Asta

Allen Institute for AI has open-sourced AstaBrief 8B, a model built to turn a research question plus retrieved excerpts into a cited report. It now sits in Asta’s Generate a report feature as Fast mode, alongside the Claude-powered Thinking mode. AI2 is also releasing the training data and an example workflow so other researchers can study the setup, run it locally, and adapt it to their own PDFs.

The pitch here is pretty specific: science users do not just want more text. They want reports that stay close to the evidence, keep the scope of claims intact, and can be checked later. AstaBrief was trained with that in mind, starting from Qwen3-8B and leaning heavily on post-training data, evaluation, and the report-generation pipeline around the model. AI2 says the aim was to see whether an open model could get close to the quality of the proprietary systems it had been using, while cutting both generation time and serving costs.

The training recipe is deliberately less fancy than the reinforcement-learning-heavy routes AI2 has explored elsewhere. Instead, it used supervised fine-tuning and direct preference optimization. That choice let the team focus on the thing that mattered most: data quality. Real user queries from Asta fed the process, and after filtering for quality, relevance, privacy, and language, AI2 ended up with 90K research-focused queries and 47K usable SFT examples. A separate DPO set came to about 6K examples, with two judge models agreeing on the winner for each pair.

The speed result is the other half of the story. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report, versus 178.5 seconds for Thinking mode. AI2 describes that as about 3.5x faster, and says the new model gets there by generating the full report in one pass rather than writing section by section and running the more expensive summarization and clustering steps used in the Claude-based path. On top of that, open weights mean institutions can run it on their own infrastructure, which matters when the research itself is sensitive or unpublished.

AI2 is careful not to oversell the evaluation. Most of the training and testing happened in 2025, and the company hasn’t rerun the full benchmark suite against today’s frontier models. But the broader lesson is clear enough: for scientific synthesis, citation density may matter more than cleverer filtering tricks, and an open model can be useful if it is built around the actual habits of researchers instead of generic chatbot behavior.