Ai2 open-sources AstaBrief, an 8B model for scientific report writing
It writes cited reports 3.5x faster than Ai2's Claude pipeline. Anyone can run it locally.
The facts
- AstaBrief 8B is built from Qwen3-8B and trained via supervised fine-tuning plus direct preference optimization (DPO), not reinforcement learning
- In Asta's Generate a report feature, Fast mode (AstaBrief) averages 51.1 seconds per report versus 178.5 seconds for Claude-powered Thinking mode, about 3.5x faster
- Training used 47K SFT examples drawn from 90K filtered real research queries, generated via a ScholarQA pipeline using Claude 3.5/3.7 Sonnet, o3, o4-mini and GPT-4.1
- DPO preference data (about 6K pairs) was built by having GPT-4.1 and DeepSeek-R1 judge competing reports, agreeing with human preference 95% of the time
- Ai2 is releasing the model weights, training data, and an example local workflow for generating reports from a researcher's own PDFs
- Evaluation used SQABench-CS2 (200 CS research questions) and DeepScholarBench (63 ArXiv-derived queries), measuring rubric score, answer precision, citation precision and citation recall
The Allen Institute for AI (Ai2) has open-sourced AstaBrief. It is an 8-billion-parameter model that turns a research question and retrieved literature excerpts into a cited report in a single pass, rather than section by section. It now runs as the 'Fast mode' inside Asta, Ai2's agentic research platform, sitting alongside a Claude-powered 'Thinking mode'. Across the full Asta pipeline, Ai2 reports Fast mode averages 51.1 seconds per report against 178.5 seconds for Thinking mode. That is about 3.5 times faster.
Ai2 started from Qwen3-8B. Rather than reach for reinforcement learning, which it calls unstable and expensive, it built the model with supervised fine-tuning and direct preference optimisation (DPO). The SFT data came from 90K real Asta user queries, filtered for quality, relevance and privacy, then used to generate 47K training reports with a mix of proprietary models: Claude 3.5 and 3.7 Sonnet, o3, o4-mini and GPT-4.1. For DPO, Ai2 built roughly 6K preference pairs by pitting Claude-generated reports against ones from o3, o4-mini, DeepSeek-V3 and DeepSeek-R1. GPT-4.1 and DeepSeek-R1 served as judges and agreed with human raters 95% of the time.
Ai2 says its early SFT runs lagged the Claude-based pipeline on answer precision and citation quality. The model wrote well but didn't stay grounded in the evidence as consistently as needed. That pushed the team to build statistics-based filters, including output-to-input token ratio and citation relevance scoring, to strip weaker synthetic examples before training the final model.
Alongside the weights, Ai2 is releasing the training data and an example workflow for generating reports from a researcher's own PDFs. Institutions can run report generation on local infrastructure. Ai2 says that matters when a research question touches sensitive or unpublished work. Ai2 notes most of this training and evaluation dates to 2025 and hasn't been rerun against today's frontier models. The comparisons should be read as evidence about its training choices, not a current state-of-the-art claim.
Why it matters
For teams building retrieval-augmented research tools, AstaBrief is a concrete data point. It shows an 8B open model, trained on filtered real-user data and judge-verified preference pairs, can approach a frontier proprietary pipeline's report quality at a fraction of the latency and presumably the serving cost. The harder lesson sits in what Ai2 admits its metrics still miss. Citation precision doesn't catch a model that cites the right paper while quietly broadening what that paper actually showed. That matters most for anyone treating these reports as working artifacts rather than disposable answers.
Primary source
Hugging Face — huggingface.coWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.