Top AI Product

Cursor Research just open-sourced the training infrastructure behind its Composer model. Mixture-of-Kittens (MoK) is a megakernel that fuses every mixture-of-experts communication and computation step into one deterministic GPU kernel — no kernel launch boundaries, no CPU-GPU sync, bitwise-identical results on every run.

One kernel instead of a pipeline

MoE layers are the bottleneck in frontier training: tokens bounce between GPUs and experts through a chain of separate kernels. MoK collapses the whole chain. The numbers: up to 2.37x MXFP8 forward throughput over the fastest public baselines — including DeepSeek’s DeepEP and NVIDIA’s own HybridEP + Megatron — plus a 1.41x end-to-end speedup in production across 512 GPUs. It already powers Composer training on tens of thousands of GPUs, built specifically for GB300 NVL72 racks. Apache-2.0, on GitHub since August 4.

Why this matters

The “Cursor is a VS Code wrapper” joke is officially dead. A company shipping deterministic megakernels for Blackwell racks is a training-infra lab that happens to sell an editor. Open-sourcing it is the DeepSeek-with-DeepEP playbook — get other labs building on your stack — except Cursor just beat DeepSeek’s own kernels at it.

You Might Also Like

• Cursor Composer 2 Takes on Anthropic and Openai With a 0 50 m Token Coding Model and the Benchmarks Back it up
• 397 Billion Parameters on a 48gb Macbook Flash moe Turns Apples 2023 Research Into Reality
• Dexter Passed 24k Stars on Github an Autonomous Agent Doing Real Equity Research
• Code Review Graph Hits 20k Github Stars an 82x Context cut for Claude Code and Cursor
• Heretic Just hit Github Trending and the ai World has Opinions