Shopify has deployed Liquid AI's non transformer architecture in production for its search function, achieving sub 20ms latency with models reportedly outperforming larger transformer models (Qwen3, Gemma3, Llama 3) while running 2 10x faster.
Mikhail Parakhin, Shopify's chief technology officer, has a one-line verdict on Liquid AI after months of live testing: it is the first non-transformer AI architecture he considers genuinely competitive with the dominant technology. He said it on a Latent Space podcast that published Wednesday, and for a company that processes roughly $200 billion in annual commerce, the implication is significant. A CTO does not stake production infrastructure on an unproven architecture.
The architecture in question is worth explaining. The AI that powers most modern language tools runs on what researchers call transformers, a design that processes language by comparing all words in a sequence simultaneously, like a librarian reading every book in a library at the same moment to understand how each relates to every other. Liquid AI takes a different approach: its models process input step by step, trading some of that simultaneous awareness for speed and efficiency. For Shopify's search function, which must return results in milliseconds without draining compute budgets, that tradeoff is the entire point.
Shopify invested $250 million in Liquid AI's Series A in December 2024, then spent the following year waiting for the technology to prove itself before the deployment switch flipped. The first production deployment, a search function that completes in under 20 milliseconds, went live in December 2025, according to Liquid AI's partnership announcement. The current flagship model, LFM-2, scores 55.23 on the MMLU benchmark and 58.3 on GSM8K — respectable but not at the frontier. On specific production-like tasks, Liquid AI's models with roughly 50 percent fewer parameters have outperformed Qwen3, Gemma3, and Llama 3 while running two to ten times faster, according to the company's blog post when the partnership was announced.
The custom generative recommender system Shopify co-developed with Liquid uses a novel architecture the company calls HSTU, which it says outperformed the prior recommendation stack in controlled testing, resulting in higher conversion rates. Liquid AI's origin traces to MIT research on liquid neural networks, a class of brain-inspired systems first described in a 2020 Nature Machine Intelligence paper. The company co-invented Mamba, a state-space model architecture that demonstrated language models could skip transformer attention machinery entirely. The original Mamba paper showed five times higher throughput than comparable transformers at inference time.
The benchmarks Liquid cites are its own benchmarks on tasks it selected. No independent third party has published comparable evaluations. The HSTU recommender claim is also unverified beyond the company's partnership announcement. Parakhin's endorsement carries weight because Shopify has skin in the game, but skin in the game is not the same as independent confirmation.
The efficiency argument will face its real test in production. If a non-transformer model can run at sub-20ms latency in a live commerce stack while matching transformer quality on the relevant tasks, the compute economics that have pushed every major lab toward larger transformers may not apply equally at the application layer, where domain-specific fine-tuning and latency constraints create room for alternatives. That is the bet Shopify made when it signed the term sheet in December 2024. Whether it was right is a question only production traffic answers.
Shopify and Liquid AI declined to comment beyond their published materials.