Meet Mako. A model built to operate the live web.
Today we're announcing Mako, a web-agent-native model purpose-built to operate the live web, trained on real task data collected from authenticated, multi-step enterprise workflows running in production.
Mako is different from other web agent models. The models powering today's AI agents were built to do, well, everything. They're expected to do everything from planning dinner to helping researchers cure cancer. These general-purpose models are extraordinary. But they weren't made to operate the web, and they definitely weren't made to operate the web at scale.
Mako was.
Mako was built to execute web tasks at scale, from reading pages and identifying the right elements to reliably completing complex, multi-step workflows thousands of times every day.
Rather than relying on brittle rules and heuristics, Mako learns what matters on a page. It maintains context across long, multi-page workflows without running into the context limits that constrain general-purpose models.
With one job to do, Mako is lean. It uses a fraction of the tokens consumed by frontier general-purpose models while dramatically reducing latency; compounding savings in both cost and time.
Mako is now the default model behind TinyFish Web Agent, and you can try it today via the API, MCP, or Playground. Just describe your goal in natural language, and let Mako do the rest.
In Playground, you can watch Mako work as it reads the page, finds what matters, completes the flow, and returns a structured result.
What makes Mako unique isn't just its architecture: it's the training data.
Mako was trained on an accumulated record of what succeeds and what fails when web agents meet the real web at scale. That experience comes from operating behind logins, through paginated and unstructured content, and across long, multi-step workflows.
Beginning with an Alibaba Qwen model which allowed the TinyFish ML team to balance power and size, Mako was trained on an accumulated record of what succeeds and what fails when web agents meet the real web at scale. That experience comes from operating behind logins, through paginated and unstructured content, and across long, multi-step workflows.
That data can't be scraped. It can't be licensed. It can only be earned by running the workflows. And Mako's capabilities come from learning from them.
Mako is available today. Try it now, read the Mako white paper, and get started at tinyfish.ai/mako.
For everything web, there's Mako.
