Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

Ai2 Open-Sources AstaBrief 8B: a Cited Research Report in 51 Seconds, 3.5x Faster Than Its Claude Pipeline

Ai2 Open-Sources AstaBrief 8B: a Cited Research Report in 51 Seconds, 3.5x Faster Than Its Claude Pipeline

The Allen Institute for AI has open-sourced AstaBrief 8B, the model behind Fast mode in its Asta research assistant. Built on Qwen3-8B and tuned with supervised fine-tuning plus preference optimisation, it writes a fully cited report in one pass, averaging 51.1 seconds across the Asta pipeline against 178.5 seconds for the Claude-powered Thinking mode. Weights are Apache 2.0, and an example workflow lets labs generate reports from their own PDFs.

The Allen Institute for AI has released the weights behind the fastest path in its scientific research assistant. On October 2 the non-profit open-sourced AstaBrief 8B, an eight-billion-parameter model that turns a research question plus retrieved literature excerpts into a single, fully cited report — the model that powers "Fast mode" in Asta's Generate a report feature.

The headline number is speed. Across the full Asta pipeline, Ai2 reports an average of 51.1 seconds per report in Fast mode against 178.5 seconds for its Claude-powered Thinking mode, roughly 3.5 times faster, with generation time itself nearly an order of magnitude lower. The gain comes from architecture rather than only from a smaller model: AstaBrief writes the entire report in one pass, skipping the snippet summarisation, clustering and section-by-section drafting that the multi-step proprietary pipeline performs.

AstaBrief is built on Qwen3-8B and tuned with supervised fine-tuning plus direct preference optimisation rather than reinforcement learning. Ai2 started from 90,000 research-focused queries drawn from real Asta user logs — filtered to remove beta-tester and bot traffic, short and non-English prompts, non-scientific requests and anything containing personal information — then generated target reports with its multi-step ScholarQA pipeline using a mix of Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1. After quality filtering, about 47,000 usable training examples remained. For the preference stage, GPT-4.1 and DeepSeek-R1 acted as judges, and only pairs on which both agreed and which matched human preferences were kept, a set Ai2 puts at 95 percent agreement.

The most effective single filter was also the simplest. Ai2 says dropping synthetic reports containing long stretches of uncited text outperformed more elaborate filtering combinations — a reminder that the bottleneck in scientific grounding is citation discipline, not model size. The model does not retrieve anything itself: retrieval and the mapping of citations to sources remain the surrounding pipeline's job, and Ai2 warns that using a prompt format different from its recommended template can degrade output.

What ships alongside the weights matters for institutions. The release includes the SFT checkpoint, the preference dataset and prompts, and an example workflow for generating reports from a local PDF corpus, so a lab can run report generation on its own infrastructure rather than sending unpublished or sensitive material to a third-party API. The model weights are Apache 2.0; the released training data and prompts carry a CC BY-NC 4.0 licence, so the collection should not be treated as wholly free for commercial use.

Ai2 is unusually explicit about the limits. It says most training and evaluation work was completed in 2025 and that it did not re-run full evaluations against 2026 frontier models, so its benchmark numbers validate an engineering approach rather than asserting a standing against current systems. Early traction is real but modest: of 374 users who tried Fast mode, 29.1 percent used it on two or more days, and 23 percent never switched back to Thinking mode.

The broader lesson is about where agentic research tooling is heading. A narrow, downloadable model that writes cited reports over local data removes a proprietary API call from the critical path and gives universities and companies a way to keep unpublished work in house. The competitive question for the frontier labs is whether general-purpose models can match the citation discipline of a small model that was trained specifically for it.

Comments (0)

Log in to join the discussion

Log In

No comments yet