@codenameposhan: Ai2 just open-sourced AstaBrief 8B, a small model that turns a research question into a cited report. You give it a question plus excerpts from retrieved papers, and it writes the whole report with citations in one pass. It's what powers the new Fast mode in Asta, Ai2's platform for science, next to the Claude-powered Thinking mode. The speed is the point: 51.1 seconds per report in Fast mode versus 178.5 seconds in Thinking mode, about 3.5x faster across the full Asta pipeline. Quality holds up in Ai2's own evals. On SQABench-CS2 it averages 87.0, between DR Tulu at 88.8 and Thinking mode at 86.2, and 90.5% of its citations support their claim. It wins 72% of LLM-judged comparisons against Thinking mode, though it was trained toward that exact ranking. The interesting part is the data: 90K real research questions, 47K reports from Claude, o3 and GPT-4.1 models, and 6K preference pairs kept only when two judges agreed. The biggest single gain came from throwing out training reports that didn't cite enough. The fine print: in a small human study, researchers preferred DR Tulu overall, and most of the evaluation was done in 2025 against that year's frontier models. It's Apache 2.0, built on Qwen3-8B, and you can run it on your own hardware, which matters for unpublished research. Sources: allenai.org/blog/astabrief and huggingface.co/allenai/AstaBrief_8B Independent briefing by CodenamePoshan, not an Ai2 post. Ai2 art and logo shown for editorial reporting. #Ai2 #AstaBrief #OpenSourceAI #AIResearch #LLM #AIEngineering

Codename Poshan

Open In TikTok:

Region: US

Saturday 03 October 2026 03:56:41 GMT