We are excited to announce MakoGenerate, available today in research preview. 

MakoGenerate is a configurable AI agent that can generate, compile, validate, and benchmark GPU code autonomously. Today, it works with CUDA and Triton, and can target Nvidia GPUs from Hopper to Blackwell. The available research preview is a playground for experimenting with different models, prompts, and hardware.

Getting Started with the Research Preview

Generating your first GPU kernel is easy and can happen in less than 60 seconds. 

• Access and setup: Create an account on generate.mako.dev

• Defining a task: Select a PyTorch reference problem from the dropdown list, add your prompt, and configure your agent

• Monitoring progress: Hit “Go” and watch your first kernel appear in 60 seconds or less

• Join the Discord server: Show off your results, trade prompts, and win prizes here https://discord.gg/WMSnXt5a 

The optional custom prompt section gives you the opportunity to add more context to the agent. You can include specific documentation, tutorials, or examples. You can also identify specific optimization techniques that you would like to see applied. Prompts can be as short as one line, or even multiple pages long.

Generating Kernels with LLMs + Search

Developing new, high-performance kernels for novel model architectures, quantization schemes, or hardware platforms remains a persistent bottleneck both in terms of engineering effort and specialist availability. 

MakoGenerate addresses these challenges by combining LLM-driven code generation with an automated feedback loop and evolutionary search. In the research preview, iterative refinement is enabled by default: the agent generates a kernel, validates it, and then performs a single iteration of improvement. While this approach is cost-sensitive, it typically does not yield peak performance on its own. To achieve state-of-the-art results, evolutionary search is used alongside iterative refinement.

When evolutionary search is enabled, the system generates multiple candidate kernels from defined inputs, then compiles, validates, and benchmarks each variant. Compiler diagnostics, occupancy estimates, and performance metrics are summarized and fed back to the LLM. Through parallel exploration and targeted mutations, such as adjusting parameters like thread-block sizes, memory access patterns, and asynchronous data movement, the search converges systematically on efficient implementations across a range of operations and architectures.

The core optimization pipeline relies on data gathered during each compile-and-test cycle. A unified harness manages multiple backends (CUDA and Triton), runs functional tests, and measures performance on target GPUs to ensure both correctness and efficiency. Feedback encoding distills compiler errors and runtime metrics into concise guidance for generating the next set of kernel candidates, steering the evolutionary search toward optimal designs.

Early Access to Evolutionary Search

We are excited to announce that we are accepting applications for the early access program to an upcoming version of the MakoGenerate agent that leverages evolutionary search. This advanced agent offers significant performance improvements and optimization capabilities. To apply for early access, please visit mako.dev and fill out the Contact Us form.