August 2026 Summaries

Cursor Router is a data-driven system that selects among price-efficient and frontier AI models based on real developer traffic, aiming to improve user satisfaction while reducing inference costs. Its two modes, Auto Intelligence and Auto Balance, use Compass, a predictor trained on subsequent user behavior to estimate task complexity, and a taxonomy that classifies tasks, domains, and modifiers to identify models with observed strengths in particular types of work. Simple turns are generally sent to a lower-cost model, while more demanding turns are assigned to eligible frontier models when measured performance gains meet a confidence threshold and fit within each mode’s cost budget. Cursor reports that Auto Intelligence now achieves above Fable-level satisfaction at 68% lower cost, while Auto Balance exceeds Opus 4.8 satisfaction at 41% lower cost, though these results are based on its internal production evaluations. The system is tested through cross-validation, held-out data, and live traffic to account for practical factors such as token use, caching, and model-switching costs, and it is intended to evolve as new models and production data become available.

Mixture-of-Kittens (MoK) is an open-sourced, highly optimized mixture-of-experts (MoE) training megakernel designed to address the bottleneck in the training of Composer, an agentic coding model, by integrating all MoE communication and computation into a single kernel for NVL72s. MoK emerged from previous efforts to enhance the MoE layer and now powers training across thousands of GPUs by achieving up to 2.37x higher throughput compared to public baselines. The kernel incorporates novel strategies such as pull-based communication to maximize NVLink bandwidth, ring token buffers to eliminate CPU-GPU synchronization, and scheduling overlaps to efficiently manage computation-communication tasks. MoK supports both BF16 and MXFP8 precision modes, with the latter offering faster performance without numerical issues. It is designed to be flexible and modifiable, allowing adaptations for various platforms, and shows significant improvements—up to 41% speedup—over previous implementations in production training stacks, thus lowering barriers to AI research and enabling more efficient model training.