Skip to content
daily-hour-news·

🛠️Ai2 Open-Sources AstaBrief, an 8B Report Model

TL;DR

Ai2 released AstaBrief, an 8B model built on Qwen3-8B that writes cited research reports in about 51 seconds. That is roughly 3.5x faster than the Claude-powered Thinking mode at 178.5 seconds.

Ai2 released AstaBrief, an 8B model built on Qwen3-8B that writes cited research reports in about 51 seconds. That is roughly 3.5x faster than the Claude-powered Thinking mode at 178.5 seconds.

Ai2 Open-Sources AstaBrief, an 8B Report Model — daily-hour-news

Key Points

1

8B parameters, fine-tuned from Qwen3-8B, open weights on Hugging Face

2

Trained on 47K SFT examples from 90K filtered queries plus 6K DPO pairs

3

51.1 s average report time vs 178.5 s for the Claude-based pipeline

4

Competitive on SQABench-CS2 rubric scores and citation precision and recall

5

Two of three researchers preferred it on citation accuracy in a small human study

Why It Matters

Citation-heavy research agents no longer need a frontier API. A small open model that is 3.5x faster changes the cost math for literature tools.

Quick Facts

Ai2AstaBriefQwen3open weightsresearch agentscitationsHugging FaceDPO

Comments

Subscribe to join the conversation...

Be the first to comment

Enjoyed this article?

Get it daily. 7am. Free. Reads in 5 minutes.

Join 3,556 builders reading daily.

Also get