Ai2 has released AstaBrief 8B, an open-weights model designed to take research questions and scientific sources and turn them into cited reports.
But unlike a normal chatbot, AstaBrief isn’t a multipurpose model. It has a much more specific job: give it a research question and relevant scientific sources, and it writes a report based on those sources with citations.
AstaBrief
Imagine you are researching a topic and you already have a bunch of scientific papers or excerpts from those papers. You could give those sources to an LLM and ask it to summarize everything, but scientific research has one pretty important problem: you need to know where the information actually came from.
Well, AstaBrief tries to solve this problem. You provide a research question + relevant scientific literature, and the model generates a structured report while citing the sources it used.
AstaBrief doesn’t search for the papers itself, though. The relevant literature has to be retrieved first and then provided to the model.
So a research system could first find relevant papers, give the relevant parts to AstaBrief, and then let the model create a cited report from them. Or you could just find the papers yourself and feed the relevant parts to the model, like in the good old days.
It’s Based on Qwen3-8B
AstaBrief uses Qwen3-8B as its base model, so it only has around 8 billion parameters instead of some massive 100B+ model.
Ai2 then trained it specifically for scientific report generation, so don’t let the relatively small parameter count fool you. The team first used supervised fine-tuning, where the model learns from examples of what a good research report should look like.
After that, they used Direct Preference Optimization. Basically, they gave the model different reports for the same research questions and showed it which ones were preferred. This helps teach the model what a better answer should look like without using a more complicated reinforcement-learning setup.
Ai2 says its final DPO training dataset contained around 6,000 preference examples.
Trained With Real Research Questions
AstaBrief’s training questions also came from real researchers, which means the data came from people who actually know their job.
Instead of only making artificial research prompts, Ai2 used real queries submitted by researchers through its scientific AI systems.
After filtering out low-quality queries, non-scientific questions, private information, and other stuff they didn’t want in the dataset, Ai2 says it ended up with around 90,000 research-focused queries.
Ai2 then used models like Claude, GPT-4.1, o3, o4-mini, DeepSeek-V3, and DeepSeek-R1 across different parts of the training process.
After more filtering, Ai2 says it had around 47,000 examples for supervised fine-tuning.
Ai2 Built AstaBrief for Speed
Ai2 also created AstaBrief with speed in mind, since its scientific research platform, Asta, already uses Claude for a more compute-heavy report generation mode.
On Ai2’s Asta research platform, AstaBrief is now used in a new Fast mode. Instead of going through multiple expensive steps to generate and organize a report, the platform gives AstaBrief the research question and retrieved literature, and the model generates the final report in one pass.
And you’re not limited to Ai2’s platform. AstaBrief has open weights, so you can download it and run it on your own hardware.
According to Ai2, the complete Asta Fast pipeline averages around 51.1 seconds per report, compared with around 178.5 seconds for its Claude-powered Thinking mode.
That makes Fast mode roughly 3.5× faster in Ai2’s testing. And obviously, waiting around 50 seconds for a research report sounds much nicer than staring at a loading animation for almost three minutes.
Quality of Results
According to Ai2’s own testing, AstaBrief performed competitively across several research-report benchmarks. They tested things like whether the report actually answered the question, whether citations supported the claims, and whether claims that needed citations actually had them.
Ai2 says AstaBrief performed competitively with its Claude-powered report pipeline and DR Tulu across several measures of answer and citation quality.
One thing to keep in mind is that Ai2 completed most of the training and evaluation in 2025, so it isn’t claiming that AstaBrief beats today’s newest frontier models.
Conclusion
So yeah, AstaBrief is basically an 8B model with one job: take scientific sources and turn them into cited reports, and according to Ai2’s testing, it does that job pretty well.
Since it’s only an 8B model, it’s much easier to work with locally than some massive research model. You can also run the model on your own infrastructure instead of sending your research to a proprietary model API.
Sources
Open-sourcing AstaBrief, the fast report-generation model in Asta
