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

Ai2 has open-sourced AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report. It is available in Asta's Generate a report feature as Fast mode, alongside the Claude-powered Thinking mode, and the model and training data are being released so others can study and build on the approach. An example workflow for generating reports from a user's own PDFs was also released.

AstaBrief started from Qwen3-8B, with most effort focused on post-training data, evaluation, and report-generation scaffolding. The team chose supervised fine-tuning (SFT) followed by direct preference optimization (DPO) instead of reinforcement learning, citing RL's instability and expense. Training used real queries submitted through the system described in the paper "Synthesizing scientific literature with retrieval-augmented LMs" and ScholarQA. After filtering for quality, relevance, and privacy, 90K research-focused queries remained; generating full-report targets with a mix of Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini, and GPT-4.1 yielded 47K usable SFT examples. The DPO set, built from a separate query subset and scored by two judge models (GPT-4.1 and DeepSeek-R1) at 95% agreement with human preferences, came to about 6K examples after filtering.

The model writes the full report in one pass rather than section by section. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report versus 178.5 seconds for Thinking mode, about 3.5× faster. Development metrics were tracked on SQABench-CS2, a set of 200 user-written computer science research questions, measuring rubric score, answer precision, citation precision, and citation recall, plus secondary evaluations on DeepScholarBench and pairwise comparisons. Most training and evaluation was completed in 2025, and the full evaluation has not been rerun against current frontier models.

Among 374 Asta users who tried Fast mode, 29.1% used it for two or more days, users generated an average of 3.67 report threads, 23% continued using it without switching back to Thinking mode, and 18% switched between modes. Fast mode received positive feedback at a rate of 84.2% versus 85.2% for Thinking mode. Ai2 said future work includes more fine-grained preference learning, RAG-plus-RL approaches, multi-turn and multi-tool capabilities, additional scientific data sources, and query decomposition.