Boston-area Liquid AI is making a very large claim in a very crowded field: Its models could eventually beat the systems built by OpenAI and Anthropic. The Cambridge startup is betting that smarter, more efficient architecture can matter as much as throwing ever-larger piles of computing power at the problem.
In a report published Monday, NBC Boston detailed Liquid AI’s push to develop foundation models in its own way, positioning the company as a local challenger to the industry’s best-known names. The pitch is not that Liquid has already unseated either rival, but that its approach could deliver comparable or better performance with fewer resources.
Liquid AI has the financial runway to take that swing. The MIT spinout raised $250 million in a venture round led by AMD, lifting its valuation to $1 billion and making it a Boston-area unicorn, according to the Boston Business Journal.
Liquid AI’s Bigger Bet Is Smaller, Faster Models
The company’s latest public work shows where that strategy is headed. Liquid AI said in February that its LFM2-24B-A2B model was designed to run across cloud infrastructure, AI PCs and mobile devices, with support from hardware and deployment partners including AMD, Intel and Qualcomm, according to the company’s model announcement.
That approach traces back to research at MIT’s Computer Science and Artificial Intelligence Laboratory, where Liquid’s founders worked on neural networks inspired in part by biological systems. The Boston Globe reported that the company’s technology differs from the transformer-heavy designs used by major AI labs, although the practical challenge remains turning an unusual architecture into a dependable commercial product.
In a July interview with The Cognitive Revolution, CEO Ramin Hasani described Liquid AI’s goal as building “device-native” intelligence that can work on phones, laptops, vehicles and other hardware outside giant data centers. The interview’s discussion also underscored the catch: Attention-based models still dominate the frontier, while Liquid’s strongest argument is efficiency on constrained devices and specialized workloads.
That leaves Liquid AI with a two-part test. It must show that its models are not merely cheaper to run, but genuinely useful enough for developers and businesses to choose them over familiar OpenAI and Anthropic tools. For a Boston startup born from MIT research, the race is no longer just about inventing a different kind of neural network; it is about proving that different can win.
