- Decision models return calibrated probabilities for all options in one call instead of generating tokens
- d1.2-ultra leads Decision Index v2.1 at 75 with a 0.8 calibration error versus 3.7
- d1.2 trails eight models at 69, and parameters, latency and weights are undisclosed
AI CURATED HOT STORY · 11/16/2026
Liquid AI's Open-Source Decision Model d1.2-ultra Tops Global Ranking
Liquid AI has released its decision models d1.2 and d1.2-ultra, marking a shift from token generation to calibrated probability outputs and topping global decision benchmarks. The d1.2-ultra ranks first on Artificial Analysis's Decision Index v2.1 private benchmark AA-Paperclip with a score of 76, three points ahead of runner-up Drex Max v1.1, while achieving a calibration error of just 0.8 versus 3.7 for competitors.
Instead of generating tokens, the model takes a state and structured question, returning calibrated probabilities for each option in a single call, and supports five answer types including value reading, span localization, bounding box, interval estimation, and multiple choice. Both models sit on the capability-cost Pareto frontier, with costs per million tasks at $133 and $63 respectively, and offer 64K context and multimodal text-image input, though only via Liquid AI's first-party API.
This breakthrough addresses the computational cost bottleneck in agent deployment. Long-chain tasks like browser operation can cost 4.8 times more than complex reasoning, making token costs a critical barrier. The open-source full-size decision model offers an alternative for the agent execution layer, potentially accelerating the transition of decision intelligence from lab to scale.
Hourly heat curve
32 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 3 reports · Latest first- Marktechpost (@Marktechpost)T2·78 pts
- 新智元 (微信公众号)T2·62 pts
- A full-size open-source model for agentic decision tasks tops a global leaderboard, outperforming Google's flagship
- Long-horizon tasks like browser operation can cost 4.8x more per run than complex reasoning, making compute the key bottleneck for agents
- The open full-scale approach offers an alternative execution layer to costly general-purpose LLMs
- Artificial Analysis (@ArtificialAnlys)T1·62 pts
- d1.2-ultra ranks #1 on the AA-Paperclip decision and extraction benchmark at 76, three points ahead of Drex Max v1.1
- It gains 4 points over d1-ultra while d1.2 improves 3 points over d1.1
- Both models sit on the capability-versus-cost Pareto frontier at $133 and $63 per million decision tasks
- 64K context, text and image input, available only via Liquid AI's first-party API