Liquid AI, an AI startup spun out from MIT, has introduced its first collection of generative AI fashions, which it refers to as Liquid Basis Fashions (LFMs).
“Our mission is to create best-in-class, clever, and environment friendly techniques at each scale – techniques designed to course of massive quantities of sequential multimodal information, to allow superior reasoning, and to realize dependable decision-making,” Liquid defined in a put up.
In accordance with Liquid, LFMs are “massive neural networks constructed with computational models deeply rooted within the principle of dynamical techniques, sign processing, and numerical linear algebra.” By comparability, LLMs are primarily based on a transformer structure, and by not utilizing that structure, LFMs are capable of have a a lot smaller reminiscence footprint than LLMs.
“That is significantly true for lengthy inputs, the place the KV cache in transformer-based LLMs grows linearly with sequence size. By effectively compressing inputs, LFMs can course of longer sequences on the identical {hardware},” Liquid wrote.
Liquid’s fashions are general-purpose and can be utilized to mannequin any kind of sequential information, like video, audio, textual content, time collection, and indicators.
In accordance with the corporate, LFMs are good at common and skilled information, arithmetic and logical reasoning, and environment friendly and efficient long-context duties.
The areas the place they fall quick at present embody zero-shot code duties, exact numerical calculations, time-sensitive data, human desire optimization methods, and “counting the r’s within the phrase ‘strawberry,’ ” the corporate stated.
At present, their predominant language is English, however additionally they have secondary multilingual capabilities in Spanish, French, German, Chinese language, Arabic, Japanese, and Korean.
The primary collection of LFMs embody three fashions:
- 1.3B mannequin designed for resource-constrained environments
- 3.1B mannequin excellent for edge deployments
- 40.3B Combination of Specialists (MoE) mannequin optimum for extra complicated duties
Liquid says it will likely be taking an open-science strategy with its analysis, and can overtly publish its findings and strategies to assist advance the AI subject, however won’t be open-sourcing the fashions themselves.
“This permits us to proceed constructing on our progress and keep our edge within the aggressive AI panorama,” Liquid wrote.
In accordance with Liquid, it’s working to optimize its fashions for NVIDIA, AMD, Qualcomm, Cerebra, and Apple {hardware}.
customers can check out the LFMs now on Liquid Playground, Lambda (Chat UI and API), and Perplexity Labs. The corporate can be working to make them out there on Cerebras Interface as nicely.
