Introducing MakoGenerate, a CUDA-writing AI agent. We're excited to make the research preview of MakoGenerate available today, completely free.

Try it for free on

You can now generate GPU kernels in #CUDA and #Triton for any arbitrary PyTorch code you have. Give it a shot!.

And this is just the beginning! There are so many new features to explore and evaluate. LLMs+Search+RL is proving to be a game changer in capability. If this kind of work excited you, apply at .

@METR_Evals Level 5 covers more complex, real-world kernels, including DeepSeek MLA, among others. MakoGenerate wins on 4/14 kernels.

KernelBench Level 2 includes slightly more complex operations with simple fusion patterns. MakoGenerate again wins on 68/100 problems.

@ScalingIntelLab KernelBench Levels 1 includes simple PyTorch operations like matmul or linear layers. MakoGenerate matches or beats torch.compile on 68/100 problems.

MakoGenerate with Evolutionary Search is already creating production-quality #CUDA kernels that beat torch.compile and expert-written kernels on real world use cases. We'll be posting examples with code throughout the week, but a few highlights are below 🧵. (ps we're hiring)

Iterative refinement is pretty neat and can yield some decent results, but the real innovation is in applying evolutionary search. Stay tuned for some cool results coming later this week.

RT @wAIeedatallah: The research preview of MakoGenerate is available today, completely free. Keep reading to see what it does, how to try i….

@nvidia

Free @nvidia Blackwell GPUs for code generation and testing???? 👀 how long can we keep this up???.

Try it for free at This research preview is a fun way to see how well different models do when it comes to GPU code generation.

We started benchmarking @Meta Llama 4 Scout on @AMD MI300X and @NVIDIA H100. Shoutout to the zero day support making life easier. Using AMD's vLLM container on long-ish context lengths (5000/1000) we get the following:.2x MI300X - 526 tps.4x MI300X - 996 tps.8x MI300X - 1144 tps

And now with the latest @AMD's AITER library, there's even more performance to be unlocked! Exciting times ahead.

The secret sauce? A combination of:. AMD's Composable Kernel for Flash Attention.GEMM tuning via PyTorch TunableOps.Liger Kernel's normalization layers (at high batch sizes).torch.compile for everything else.

There's no single "best" kernel library for operations like attention. The optimal choice depends on your specific workload, batch size, and hardware. Our Mako Compiler automates the process of finding the best combinations.

Kernels Together Strong 🦧 Our last blog post showed how combining multiple kernel libraries can deliver state-of-the-art performance for AI models. We achieved up to 60% speedup on the FLUX.1-schnell model using #AMD MI300X GPUs! #AI #GPUOptimization