Ai2 AstaBrief 8B

Overview

AstaBrief is an 8-billion-parameter scientific report generation model that Ai2 open-sourced on October 2, 2026. It was post-trained from Qwen3-8B using SFT followed by DPO, without reinforcement learning. The job it does is narrow: given a research question and a batch of retrieved literature excerpts, it produces a cited report in one pass.

It is already live in Ai2's Asta product as Fast mode inside the Generate a report feature, sitting alongside the Claude-powered Thinking mode. Speed is the most immediate difference. AstaBrief is trained to generate the final report directly from the query and retrieved excerpts in a single pass, skipping the expensive snippet summarization and clustering stages of Thinking mode and avoiding section-by-section generation. Across the full pipeline it averages 51.1 seconds per report against 178.5 seconds for Thinking mode, roughly 3.5x faster.

Weights are on Hugging Face at allenai/AstaBrief_8B and the training data is released alongside them. The official blog does not name a specific license, only describing the model as open-sourced with open weights.

Key Features

• Cited reports in a single pass: Takes a research question and retrieved literature excerpts and outputs a complete report directly rather than generating it section by section, with citations attached to the claims they support
• Live in Asta as Fast mode: Available in Asta's Generate a report feature alongside the Claude-powered Thinking mode, averaging 51.1 seconds per report against 178.5 seconds for Thinking mode
• Weights and training data released together: Model weights live at allenai/AstaBrief_8B on Hugging Face, with training data published alongside: roughly 47K SFT samples and about 6K DPO preference pairs
• Training filtered for citation quality: Synthetic reports were filtered on output-to-input token ratio, citation relevance, citation density and citation diversity, with the largest gains coming from dropping low citation-density samples
• Preference data gated by two agreeing judges: DPO pairs were compared by two judges, GPT-4.1 and DeepSeek-R1, and kept only when both agreed; the judges matched human preferences 95% of the time
• Self-hostable: Open weights let institutions deploy the model on their own infrastructure, including behind a firewall, with a sample workflow for generating reports from your own PDFs

Use Cases

• Institutions and research groups that need literature synthesis running behind a firewall or on their own infrastructure
• Research workflows that want a cited first draft and then have a researcher refine it
• Literature reviews that process many research questions at once and care about per-report latency
• Teams studying citation attribution and grounding, who can build on the released training data

Pros

• Runs at 8B parameters, so a single professional or consumer GPU can carry it
• About 51.1 seconds per report end to end, roughly 3.5x faster than the Claude-powered Thinking mode
• Both weights and training data are open, so you can self-host it or keep researching on top of it
• Training data comes from real research queries in Asta and ScholarQA, 90K after filtering
• In a small 14-question human study, two of three researchers preferred it on citation accuracy
• The team states plainly that training and evaluation were completed in 2025 and were not rerun against the newest frontier models

Pricing

Model weights and training data are free to download, and the official blog does not name a specific license. Inside Ai2's Asta product it is offered as Fast mode alongside the Claude-powered Thinking mode and can be used directly at asta.allen.ai.

Summary

The value of AstaBrief is not that it writes better reports than frontier models. Ai2's own conclusion is only that it is competitive with, or within range of, the Claude-powered pipeline and DR Tulu across several measures of answer and citation quality, and in the small 14-question human study DR Tulu still led on overall preference. The value is that it collapses the cost of producing a cited report: 8B parameters, 51 seconds, open weights, runnable on your own machine.

Citation accuracy is where it genuinely holds up, with two of three researchers preferring it on that axis. For a tool that only accelerates the literature synthesis step, that trade-off is clear. One caveat worth noting: the team says training and evaluation were largely done in 2025 and were not rerun against the current frontier models.

Version History

• AstaBrief 8B open-sourced (2026-10-02): Ai2 open-sourced AstaBrief, an 8-billion-parameter scientific report generation model post-trained from Qwen3-8B with SFT and DPO, turning a research question and retrieved literature excerpts into a cited report in one pass. It is live in Asta's Generate a report feature as Fast mode, averaging 51.1 seconds per report against 178.5 seconds for the Claude-powered Thinking mode. Weights are published at allenai/AstaBrief_8B on Hugging Face with training data released alongside, about 47K SFT samples and 6K DPO preference pairs, drawn from 90K filtered research queries taken from real Asta and ScholarQA user logs