Ai2 released AstaBrief on 2 October 2026, an 8B open-weights model that turns a research question and retrieved literature excerpts into a cited scientific report. In Ai2’s Asta platform its Fast mode averages 51.1 seconds per report against 178.5 seconds for the Claude-powered Thinking mode, about 3.5 times faster across the pipeline. Ai2 says most of the training and evaluation was done in 2025 and has not been rerun against today’s frontier models.
Ai2 says it built AstaBrief 8B starting from Qwen3-8B. It is available in Asta’s Generate a report feature as Fast mode, alongside the Claude-powered Thinking mode, and Ai2 is open-sourcing the model and the training data.
Ai2 also says it is releasing an example workflow for creating reports from a researcher’s own PDFs. It says open weights let institutions run the model on their own infrastructure, which it calls “necessary when research questions reveal sensitive or unpublished work.”
From Ai2’s post: across the full Asta pipeline, Fast mode (AstaBrief 8B) averages 51.1 seconds per report against 178.5 seconds for the Claude-powered Thinking mode, which Ai2 calls about 3.5 times faster. Ai2 says most of the evaluation was done in 2025.
Business Pill · FEWER, BETTER EXAMPLES
A one-minute explainer of the idea behind this story: data quality. It teaches the concept, not this story’s figures.
The key insight: Ai2 credits its strongest training gain to one plain data filter, keeping reports that cite their claims, and says more elaborate filtering did not help. Its speed figure compares two pipeline set-ups, not two models alone.
The Speed Claim
Ai2 says the result is “nearly an order-of-magnitude reduction in report generation time compared to the proprietary models we tracked.” In the same sentence it gives the pipeline-level figure: Fast mode averages 51.1 seconds per report against 178.5 seconds for Thinking mode, which it calls about 3.5 times faster.
Ai2 says the speed comes partly from a change in method. It trained AstaBrief to write the final report in one pass, bypassing the snippet summarization and clustering stages that Thinking mode uses, and not writing the answer section by section. It says this was possible “without sacrificing performance.”
How It Was Trained
Ai2 says it started from real queries submitted through its ScholarQA system and filtered them to a pool of 90K research-focused queries. For supervised fine-tuning it generated report targets with its multi-step pipeline, backed by a mix of proprietary systems: Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1. After quality filtering that left 47K usable examples.
For preference training it built about 6K pairs, with two judge models, GPT-4.1 and DeepSeek-R1, and kept only pairs where both agreed. Ai2 says it chose supervised fine-tuning plus direct preference optimization over reinforcement learning because RL-based training “can be unstable and expensive.”
The Filter That Mattered
Ai2 says it tested four statistics-based filters on the synthetic training reports, and that the strongest gains came from removing reports with low citation density, meaning a low share of statements with at least one citation.
Ai2 says that was among the clearest lessons of the project: “more elaborate filtering wasn’t necessarily better.” It says a simple signal, whether the synthetic reports consistently cited their claims, was more useful than several more complicated combinations it tried.
Quality and Early Usage
Ai2 says AstaBrief was competitive with the Claude-powered pipeline and with its DR Tulu system across several measures of answer and citation quality. In a separate 14-question human study, with three scientific researchers each contributing 4 to 5 questions, Ai2 says DR-Tulu wins on overall preference, while two of the three researchers prefer AstaBrief on citation accuracy.
On usage, Ai2 says that among 374 Asta users who tried Fast mode, 29.1% used it on two or more days and users on average generated 3.67 report threads with it. It says 23% continued and never switched back to Thinking mode, and another 18% switched between the two. Fast mode received positive feedback at 84.2% against 85.2% for Thinking mode, and Ai2 says the feedback is too sparse to draw strong conclusions.
The Structural Read
The post bundles three changes that are easy to read as one. The model is smaller, at 8B. The pipeline is shorter: Ai2 says AstaBrief writes the final report in one pass, bypassing the snippet summarization and clustering stages that Thinking mode uses. And the training data was filtered for citation density. The 51.1 against 178.5 seconds is a figure for the full Asta pipeline, and the text gives no split of how much of the gap comes from each change.
The model’s training data came from the systems it sits beside. Ai2 says the supervised targets were generated with a mix of Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1. It says each preference pair set a ScholarQA report, typically backed by Claude 3.5 or 3.7 Sonnet, against one from o3, o4-mini, DeepSeek-V3 or DeepSeek-R1, with GPT-4.1 and DeepSeek-R1 as the judges. Thinking mode, the comparison point for the speed figure, is Claude-powered.
Ai2 also fences its own evidence. It says the training and evaluation were mostly completed in 2025 and the full evaluation has not been rerun against today’s frontier models, that the human study covered 14 questions and three researchers, and that Fast mode feedback is too sparse for strong conclusions.
Ai2 — Open-sourcing AstaBrief, 2 October 2026
“That was one of the clearest lessons from the project: more elaborate filtering wasn’t necessarily better.”
Three Implications
A SPLIT PRODUCT Asta now offers both: Fast mode on Ai2’s open model and Thinking mode on Claude. Ai2 says 23% of users who tried Fast mode never switched back to Thinking mode and another 18% moved between the two, so in this product the choice of model is a setting the user makes, not a fixed backend.
WHERE THE EFFORT WENT Ai2 says it spent much of the project working out how to generate, select and filter training examples, and chose supervised fine-tuning and preference optimization over reinforcement learning because RL-based training can be unstable and expensive. In Ai2’s account the work sat in the data and the evaluation, not in a more complex optimization method.
WHAT REMAINS UNKNOWN The post gives no dollar serving cost, no split of the speed gain between model and pipeline, and no independent replication. Its evaluation mostly dates from 2025 and has not been rerun against current frontier models, and this publication did not read the post’s chart images.
What Is Not Established
Ai2 says the numbers are best read as validation of its engineering approach at the time, not as a claim about where the base model sits against today’s frontier. It says it has not rerun the full evaluation against current frontier models.
The text of the post gives no dollar figure for serving costs, although Ai2 says the aim included reducing them. This publication did not read the post’s charts as images, did not run the model and did not seek a response from Ai2.
Business Engineer Framework
The Map of AI — Where Open and Frontier Models Sit in the Stack
The Map of AI places more than 200 companies across nine layers of the stack, from silicon to application. Whether a product’s model comes from an open or a closed layer, and who holds the leverage there, is the frame a release like this invites.
Ai2 says an 8B open-weights model, trained on filtered real queries and outputs from proprietary systems, writes cited scientific reports about 3.5 times faster than the Claude-powered mode in its own Asta pipeline, at 51.1 seconds against 178.5. It says the biggest data gain came from keeping only reports that cite their claims, and that its evaluation mostly dates from 2025.
A note on sourcing. We drew entirely on Ai2’s blog post of 2 October 2026, reading the text only; we did not examine its chart images, run the model, or reach out to Ai2 for comment. Every result and figure above is Ai2’s own claim about its own system, and we have not checked any of it independently. Nothing here is a forecast or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
