TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Ai2 has open-sourced AstaBrief 8B, a model designed to generate cited scientific reports from research questions and retrieved literature. In Asta, its Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for the Claude-powered Thinking mode; Ai2 says the evaluation was completed mostly in 2025 and has not been rerun against current frontier models.
Ai2 has open-sourced AstaBrief 8B, a model built to turn a research question and retrieved scientific literature excerpts into a cited report. The model is available in Asta’s report-generation feature as Fast mode, and Ai2 is also releasing training data and an example workflow intended to help others run or adapt report generation on their own infrastructure.
Ai2 reports that Asta’s Fast mode averaged 51.1 seconds per report across its full pipeline, compared with 178.5 seconds for the Claude-powered Thinking mode. That is about 3.5 times faster in the comparison described by Ai2. The figures refer to the Asta system’s report generation, not a standalone model speed benchmark.
AstaBrief starts from Qwen3-8B. Ai2 says it focused on post-training data, evaluation and the surrounding report-generation system, using supervised fine-tuning and direct preference optimization rather than reinforcement-learning-based training. The model is trained to write a complete report in one pass from a query and retrieved excerpts, bypassing parts of the longer pipeline used by Thinking mode, including snippet summarization and clustering.
The release includes model weights and training data, as well as an example workflow for creating reports from researchers’ own PDFs. Ai2 says the goal is to let institutions and other users study the approach, reproduce parts of it and adapt report generation to local needs. A downloadable model can also support deployments on an institution’s own infrastructure, which Ai2 identifies as relevant when research questions involve sensitive or unpublished work.
Faster Reports With Open Weights
The release offers researchers an alternative to relying solely on a proprietary model or hosted service for literature synthesis. Open weights can give institutions more control over where report generation runs, while the released data and workflow make it possible for other teams to inspect and build on Ai2’s approach. This may be useful where unpublished research or other sensitive material makes external processing difficult, though the announcement does not establish that every local deployment will meet an institution’s privacy or security requirements.
The reported speed difference may also matter in iterative research workflows: a faster draft can give users a starting point to review and refine. But speed is only one part of scientific usefulness. Reports must accurately represent the evidence, maintain appropriate citations and avoid extending findings beyond what the cited studies support. Ai2 says those were central goals of its training and evaluation; the published timing figures alone do not establish how well the model performs across all disciplines or research tasks.
By releasing the model and training materials together, Ai2 is making its report-generation recipe more available for scrutiny than a closed service would be. Independent testing will be needed to establish how its quality, citation behavior and practical costs compare with other models in different settings.
scientific report generation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How AstaBrief Was Built
Asta is Ai2’s agentic platform for scientific work, and its Generate a report feature uses retrieved literature to address research questions. Ai2 describes users asking questions with multiple constraints, such as comparing approaches across a body of literature while accounting for a particular method, population or setting. The resulting reports can serve as research artifacts users revisit, rather than one-off answers.
Ai2 says the training process began with real user queries submitted through systems associated with its research on synthesizing scientific literature and the ScholarQA framework. The team emphasized the selection and filtering of training examples, including citation-focused filtering and preference data. The stated design aim was to improve relevance, structure, evidence grounding and long-form report quality without relying on a more complex reinforcement-learning setup.
The announcement situates the work within Ai2’s broader effort to develop open models for scientific discovery, including research connected to the U.S. National Science Foundation’s OMAI initiative. Ai2 says it chose a comparatively simple training recipe partly because reinforcement-learning methods can be unstable and costly, and it wanted an approach that would be easier to operate, debug and revise.
“Fast mode averages 51.1 seconds per report compared with 178.5 seconds for Thinking mode, about 3.5× faster.”
— Ai2, in its Hugging Face announcement
research literature summarization tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Current Quality Comparisons Remain Open
Ai2 says most of the training and evaluation described were completed in 2025, and that the proprietary models used for training data and comparison reflected the frontier at that time. The team has not rerun the full evaluation against current frontier models, so the release does not establish how AstaBrief compares with today’s leading systems.
The announcement gives aggregate report-generation times but does not provide, in the supplied material, enough detail to determine performance across particular scientific fields, query types or institution-specific deployments. It also does not establish that citations are correct in every generated report or that the model consistently preserves the limits of the underlying evidence. Those questions require review of the evaluation details and independent testing.
Ai2 describes the model as open-sourced and says it is releasing weights, training data and an example workflow. The supplied announcement does not specify all licensing terms, deployment requirements or operational costs for running the model locally. Those details will affect how easily institutions can adopt it.
AI-powered scientific report generator
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing the Release in Research Workflows
The immediate next step is for researchers and developers to examine the released weights, training data and PDF workflow, then test them against their own literature and reporting requirements. Teams evaluating the model can check whether its citations support the claims made, whether reports preserve study limitations and how output quality changes across fields and question types.
Ai2 says the example workflow is a starting point for local report generation and presents AstaBrief as a test case for open models tailored to scientific work. The announcement does not set out a date for a refreshed comparison with current frontier systems. Until one is published, the reported speed and quality results should be read as findings from the specific system and evaluation Ai2 describes, not as a current general ranking.
local deployment AI research tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is AstaBrief?
AstaBrief 8B is an Ai2 model designed to generate cited scientific reports from a user’s research question and retrieved literature excerpts. It is offered in Asta’s Generate a report feature as Fast mode.
How much faster is AstaBrief in Asta?
Ai2 reports an average of 51.1 seconds per report for Fast mode and 178.5 seconds for Claude-powered Thinking mode across the Asta pipeline. The company characterizes this as about 3.5 times faster.
What has Ai2 released?
Ai2 says it is releasing AstaBrief’s model weights and training data, along with an example workflow researchers can adapt to generate reports from their own PDFs.
Has AstaBrief been compared with current leading models?
Not in a full updated evaluation described in the announcement. Ai2 says most of its training and evaluation work was completed in 2025 and that it has not rerun the full comparison against today’s frontier models.
Can institutions run AstaBrief locally?
Ai2 says the open weights allow institutions to run the model on their own infrastructure, and it provides an example workflow for local report generation. The announcement does not specify every deployment requirement or confirm that local operation automatically satisfies an institution’s privacy, security or compliance needs.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
