PALO ALTO, CA, July 22, 2026 (GLOBE NEWSWIRE) -- TinyFish today announced Mako, the world's first web-native AI model trained on production enterprise web task data. Purpose-built to operate the live web, Mako learns from running authenticated, multi-step workflows for enterprises in production.
Enterprises are putting AI agents to work on the web, and the models doing that work were built for something else. General-purpose models are extraordinary reasoners, but operating a web page is an execution job: reading the page, finding the element, completing the flow, reliably and at scale. Mako was built for exactly that job.
Mako reads web pages natively instead of re-deriving their structure from raw text, learns what on a page matters instead of relying on hand-written rules, and holds long multi-page workflows without the context limits that constrain general-purpose models. Mako encodes web elements and their relationships directly into the model and caches page state across a task, so each step costs what the step requires, not what the whole task history weighs.
The efficiency claim has a cleared number behind it: Mako's current candidate runs 35 billion total parameters with 3 billion active at inference, against frontier models estimated in the trillions.
What makes Mako difficult to replicate is not the architecture alone. It is the training data: the accumulated record of what works and what fails when agents meet the real web at scale, collected across the hardest-to-reach parts of the web, behind logins and across multi-step flows. That data cannot be scraped or licensed. It comes only from running the workloads.
"AI has read nearly everything humanity has written, and the returns on reading are running out. The next leap comes from experience, from systems that act in the real world and learn from what actually happens. Mako is built for exactly that, on the largest environment for real action there is: the live web," said Sudheesh Nair, CEO and co-founder of TinyFish.
Mako runs inside the TinyFish platform, where enterprises pair the reasoning model of their choice with a model built for the web work. The reasoning model owns understanding the goal and planning the task; Mako owns the doing.
"Testing at the speed of AI-generated code means driving real browsers reliably, thousands of times a day. A web-native model is the right tool for that, and the economics follow," said Yunhao Jiao, CEO of TestSprite.
Mako is generally available to enterprises starting July 21. Businesses can see Mako complete a real workflow, read the Mako white paper, and get started at https://tinyurl.com/4czs275w.
About TinyFish
TinyFish builds the web operating layer for enterprise AI: search, retrieval, and autonomous operation across the live web, including the authenticated, multi-step workflows that read-only tools cannot reach. Founded by Sudheesh Nair (former President of Nutanix), Shuhao Zhang (former Engineering Leader at Meta), and Keith Zhai (former The Wall Street Journal senior correspondent), TinyFish serves enterprises worldwide from its headquarters in Silicon Valley. To learn more, visit tinyfish.ai.
Sky Zhang, Growth at TinyFish skyzhang@tinyfish.ai https://www.tinyfish.ai/
