About the company
FrontierRL is building Star, the
ultimate LLM for real-time throughput.
We are an MIT startup training and serving our own state-of-the-art model. Video, audio, and live text streams can now all be processed in real time, on hardware everyone else overlooks.
Why we're different
Other people are optimizing on GPUs. We went back to the fundamentals.
The rest of the field trains the biggest model it can afford, then fights physics to make it fast. We started from the other end and rebuilt three things everyone else treats as settled:
01 · Model size
Unique size
Star is sized for the work, not for the leaderboard. That's why it decodes thousands of tokens per second on workstation GPUs. That speed turns agents from a demo into a tool you can afford to run all day.
02 · Training
Unique training
Built from years of expertise in LLM architecture and distillation. Star wasn't trained like other models, and it doesn't behave like them. She can process full speed, real-time video, audio, and text stream for pennies.
03 · Communication
Unique agents
Star agents don't just run alone. They coordinate. Multiple agents fighting together in a game, or writing up a codebase, communicate in ways other agent stacks simply aren't made for.
The FrontierRL design philosophy
Coming from a neuroscience background, our agents are built like
neurons in a brain, not electrons that circle each other concurrently.
What we make
Star is our own model. Not a wrapper, not a reseller. It generates thousands of tokens per second, as measured on real OSWorld tasks on workstation GPUs. That speed, on inexpensive hardware, is what makes our agents practical at a fraction of typical model costs.
Star agents can click and type across any app, platform, or API the same way a person does. Our public demos show Star designing presentations in PowerPoint, rebuilding a racecar in CAD, sourcing candidates on LinkedIn, and playing complex real-time games, all on real software.
How teams use it
Currently we are using Star through a Hermes harness for the purposes of benchmarking and proof of concept. API keys are available upon request until we launch on OpenRouter.
We are building an interface for non-developers to train their own Star agents; users would share their screen and show complex workflows across GUIs, talk through the steps they can't show, and drop in the policies, guides, and examples their team already uses. Our Workflow Miner turns that into protocols Star agents can follow. Agents would run hands-free, with approvals and a clear record when you need them. Consider piloting with us to learn more!
Who we are
FrontierRL was built by engineers from MIT, Amazon, and Immunai, spanning machine learning, large-scale systems, and computational biology. We participate in MIT TNT Research and the Together.ai Accelerator.
Contact
Email us at [email protected], or book an intro and show us one messy, high-value workflow your team repeats regularly.
FrontierRL · https://www.frontierrl.com