Performance at every scale and in every domain: how and where Liquid AI delivers ROI
Thought piece

Performance at every scale and in every domain: how and where Liquid AI delivers ROI

Quality, Efficiency, Explainability: What’s different about our models 

On average, our models deliver both quality and efficiency, provide transparency into how and why the model is generating its outputs, and consume significantly less power and memory than their equivalent transformer counterparts doing the same task. 

A closer look: 

  • Lower power and memory footprint: Other foundation models require a huge amount of memory on-device, necessitating more complex hardware, and limiting use-cases for on-device compute. With LFMs, it’s possible to fit up to 1 million tokens-worth of data and map it onto 16 gigabytes of memory while still achieving high-quality outputs.  
  • Unparalleled transparency: It’s often difficult to understand how transformer-based models generate their outputs. Our DevKit solves for that pain point. By simply typing “model.explain,” the model explains all the axes it used to generate its output – providing organizations with a clear, step-by-step process of how their model arrived at its answer.  
  • Achieve efficiency and quality: We take a holistic approach to model design – integrating data, training algorithms, and post-training procedures – allowing us to optimize how each component interacts with the whole. This efficient approach, combined with our fine-grained internal evaluations, enables us to identify the strengths and weaknesses of various architectural, data, and training approaches, ensuring continuous improvement and exceptional performance. 

Speed to Market Matters: Deploy in Weeks, not Months 

Companies can solve for their questions of transparency and explainability, energy and memory usage, latency issues, and privacy concerns – and they can spin it all up in days instead of weeks or months. 

With greater capabilities, explainable outputs, and lower footprints, LFMs meet organizations where they are – enabling them to reach their AI ambitions across modalities and industries with their current resources, people, and technology stack. 

Models that Scale: What you’re buying 

Liquid offers vertical-specific base LFMs available in three different scales capable of operating on a Raspberry Pi, mobile devices, and at the enterprise level. Each model is fine-tuned by Liquid’s internal evaluations, and, in the future, can be further fine-tuned on-prem or in the cloud by organizations to meet their unique development and deployment scenarios. 

  • 1.3B model: Suitable for deployment in highly resource-constrained environments where the model needs to operate on low-power, small-scale devices like a Raspberry Pi. Watch a 1.3B model ported to a Raspberry Pi respond to “explain the isotopes about Hydrogen.” 
  • 3.1B model: Optimized for edge deployments on devices such as phones, laptops, robots, or wherever the compute takes place. Watch a 3.1B speech model have a conversation about things to do in Boston and provide a tip about networking at a tech launch event, then check out an offline 3.1B text model share Halloween costume ideas, and lastly – watch a tailored 3.1B model operating on a custom application take voice input from a tech support agent and turn it into structured outputs. 
  • 40.3B Mixture of Experts (MoE) model: Designed for deployment in the cloud or on-premises using a single GPU. It is the most capable and largest variant, making it ideal for more complex tasks. Watch a 40.3B model provide a response using additional data from a web search that wasn’t originally in its training data.  

How and Where LFMs are Applied 

These LFMs not only deploy where other models can’t, they can understand a variety of sequences. 

  • Language LFMs: Handle a wide array of natural language processing tasks, from chatbots and customer support automation to complex reasoning and mathematical queries. 
  • Bio LFMs: Analyze biological data to learn the language of proteins and generate new ones, accelerating biotechnology and therapeutic research. 
  • Drive LFMs: Simulate the physical world to train and stress-test autonomous systems, predicting outcomes regardless of time, driving conditions, or weather. 
  • Transaction LFMs: Examine financial transactions to detect fraud, safeguarding finances by predicting and identifying suspicious activities. 
  • TimeSeries LFMs: Are ideal for complex, structured time series data for forecasting, trend analysis and monitoring, and handling non-linear relationships more effectively than traditional models. 

Where you can test LFMs and learn more 

Currently, organizations can test Liquid’s 1.3B and 3.1B models through the cloud on our Liquid AI Playground, and select organizations can join us in co-development. Models for on-prem or on-device deployment are still in alpha, and we can contact you when they are ready. 

Tune into our recent launch event to learn more or get in contact with our team directly here to start building your AI solution.