Allen AI open-sources AstaBrief 8B, a scientific report model that writes a cited paper in 51 seconds: Reveneau

Why it mattersA team that has been paying for Claude as the writing step in a research assistant can swap in an 8B model running on its own hardware, pay nothing per report, and keep sensitive research off a vendor's server.

A free, open-weight model now writes cited research reports in a third of the time a Claude pipeline takes.

The Allen Institute for AI published AstaBrief 8B on Hugging Face on 2 October 2026, with a blog post by Kyle Wiggers of Ai2 Comms. The model is fine-tuned from Qwen3-8B, generates a full report in one pass instead of section by section, and is released as open weights at allenai/AstaBrief_8B. It is the writing step inside Asta, Ai2's scientific research assistant, under the "Fast mode" label.

The number the post leads with

Ai2 reports a measured 51.1 seconds per report on average for AstaBrief, against 178.5 seconds for the Claude-based "Thinking mode" pipeline the Fast mode replaces inside Asta. That is 3.5 times faster on the same task, measured on Ai2's own harness. The paper's authors are careful to note that the comparison was done on 2025-era frontier models and has not been re-validated against today's frontier.

On SQABench-CS2, a 200-question computer science benchmark, AstaBrief scores roughly 82 percent on the rubric score, below the Claude pipeline at about 85 percent and DR Tulu at about 87 percent. It is ahead on citation precision at about 82 percent versus the Claude pipeline at about 80 percent and DR Tulu at about 76 percent. These are the authors' own numbers.

How the training set was built

The post describes the training data as 47,000 supervised fine-tuning examples, filtered down from 90,000 real queries that Asta users ran, and about 6,000 preference pairs for the preference-tuning step. The supervised examples were written by Claude 3.5 and 3.7 Sonnet, o3, o4-mini, and GPT-4.1. The preference pairs were judged by GPT-4.1 and DeepSeek-R1, with both judges required to agree, which the authors say matched human preference 95 percent of the time.

Ai2 writes that it tried stronger preference-tuning methods and reinforcement learning first, and settled on direct preference optimization because it was cheaper, simpler and more stable. The data filter that mattered most was citation density: keeping only training examples whose citations were dense enough to teach the model to ground its claims.

Who it is for in practice

A university or company lab that was paying Claude per research report, or that could not send its queries and documents out at all, can now run the writing step on its own hardware. The weights are open, the Hugging Face page documents the model format, and the blog links an example local workflow at allenai/ai2-scholarqa-lib.

Ai2 shares usage data from Asta that is a mild endorsement, not a measurement: of 374 users who tried Fast mode, 29.1 percent used it for two or more days and 23 percent went on to use it only, versus 18 percent who kept switching between Fast and Thinking modes. In a 14-question human study by three researchers, two of the three ranked AstaBrief first on citation accuracy, and the user satisfaction rate came in at 84.2 percent versus 85.2 percent for Thinking mode.

The post does not state a licence on the model itself and no benchmark figure has been independently reproduced, so a buyer evaluating this should run their own task on their own documents before swapping it in.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.