Liquid AI Introduces the Groundbreaking d1 Series Decision Models
On October 7, Liquid AI published an exciting new blog post announcing a major technological leap. The company officially launched two powerful open-weight decision models under the new d1 series. You can view the official announcement on their social media update. Firstly, the remarkable d1-3B model supports massive 3.12 billion parameter text and image analysis. Meanwhile, the highly efficient d1-omni-600M utilizes 587 million parameters to process text, image, and voice inputs simultaneously.
Entering the Highly Competitive AI Decision Arena
The broader artificial intelligence industry witnessed explosive interest in decision models following the phenomenal debut of the Jev model in September. Subsequently, major industry giants like OpenAI began launching their own similar analytical systems. Now, Liquid AI has boldly entered this intensely competitive battlefield as a formidable new challenger. Furthermore, the newly unveiled d1-3B and d1-omni-600M models are already publicly available on Hugging Face. Therefore, eager users can immediately download, meticulously fine-tune, and deploy these cutting-edge tools.
Deep Dive into the Powerful d1-3B Model
The impressive d1-3B boasts an astonishing 3.12 billion parameters within its complex architecture. The development team trained it extensively on the sophisticated LFM2.5-VL-3B visual language model. Consequently, it effortlessly handles complex combinations of text and image inputs. During rigorous evaluations on the Decision Index v0.2.1 public test set, this model achieved a stellar score of 48.57 points. Liquid AI confidently asserts that this remarkable performance significantly outpaces all other competing models currently operating under 10 billion parameters.
Exploring the Versatile d1-omni-600M
Conversely, the remarkably agile d1-omni-600M operates efficiently with 587 million carefully tuned parameters. Developers built this system fundamentally upon the robust LFM2.5-Encoder-350M bidirectional encoder. This highly versatile model smoothly supports combinations of text and image inputs, or alternatively, text and voice pairings. Notably, this release represents Liquid AI’s first official experimental checkpoint for a truly multimodal decision model.
Exceptional Inference Speed and Performance
When evaluating sheer inference speed, the d1-3B demonstrates truly blazing-fast capabilities. Running on a powerful NVIDIA RTX 4090, it remarkably answers a single complex question in merely 8 milliseconds. Additionally, processing a detailed 384-pixel image requires only 102 milliseconds. Furthermore, when deployed on a Jetson AGX Thor, a single query takes a swift 16 milliseconds. Similarly, operating on a Jetson Orin Nano yields an impressive response time of exactly 50 milliseconds.
Real-World User Testing Results
Independent user testing has enthusiastically confirmed these remarkable performance metrics across various hardware setups. For instance, testing the d1-3B on a standard GeForce RTX 3060 equipped with 12GB of video memory yielded excellent results. A single analytical judgment took only 25 milliseconds to complete. Meanwhile, processing a standard 640×480 pixel image required just 146 milliseconds. Crucially, the peak video memory consumption hovered efficiently around 6GB. Furthermore, other dedicated users reported that the model consistently achieved lightning-fast 8-millisecond single judgments when utilizing an RTX 3090.
