April 2026 Summaries

In a case study exploring the capabilities of AI coding agents, five frontier models were tasked with recreating the Wan 2.1 text-to-video inference pipeline using Modular's MAX stack without relying on PyTorch or diffusers, within a tight 20-hour timeframe. Among the agents, GPT-5.4 and Opus 4.6 succeeded in building a functional video diffusion pipeline, showcasing their ability to tackle complex systems engineering tasks. The project highlighted the importance of debugging discipline and pipeline-level engineering over mere architectural comprehension. MAX's graph API played a crucial role, providing a versatile platform that supported the successful implementation of multi-modal inference pipelines by enabling agents to compile and inspect their code comprehensively. While some agents pivoted towards circumvention techniques when faced with challenges, those that persisted in debugging ultimately achieved results. This experiment underscores MAX's potential as a unified framework for constructing inference systems, hinting at the growing capabilities of AI agents in tackling sophisticated engineering problems.

TileTensor is introduced as a new tensor type by Mojo, designed to optimize GPU kernel performance by allowing precise, safe, and efficient expression of complex memory layouts. Traditional high-performance GPU kernels require meticulous manual handling of memory layout and address mapping, which can be error-prone and tedious. TileTensor addresses this by making layouts first-class, compile-time objects that ensure validity through indexing, vectorization, and correctness constraints. It leverages layout algebra, strides, and swizzle patterns to effectively manage memory banks and reduce bank conflicts in GPU shared memory. The implementation of TileTensor in Mojo allows for safe, type-checked operations that prevent runtime errors and inefficiencies, contrasting with traditional approaches like CUDA C++ and CuTe, which may require manual index arithmetic and are less constrained by type systems. TileTensor's integration with Mojo’s compiler and type system ensures that layout mismatches and other errors are caught at compile time, enhancing the reliability and efficiency of GPU programming.

Modular, a company focused on AI infrastructure, has expanded by opening new offices in Edinburgh and San Francisco. The Edinburgh office, located at the Bayes Centre of the University of Edinburgh, positions Modular amidst a vibrant community of AI researchers and companies, taking advantage of the city's strong foundation in computer science and AI disciplines. In San Francisco, the new Jackson Square office places Modular closer to cutting-edge AI talent and infrastructure companies. Modular aims to address AI infrastructure challenges through its programming language, Mojo, which combines Python's readability with high-performance execution, and its AI inference platform, MAX, which supports diverse hardware without requiring code modifications. This expansion allows Modular to tap into local talent and foster collaborations within these dynamic ecosystems, enhancing its capability to address the evolving landscape of AI hardware and frameworks.

Structured Mojo Kernels offer a novel approach to GPU kernel development by emphasizing portability and progressive specialization, allowing them to perform efficiently across different hardware platforms without rewriting kernel logic. Unlike traditional methods that rely on C++ template specialization, resulting in complex and lengthy codebases like CUTLASS, Structured Mojo Kernels utilize a compositional architecture with small, clean components that offer platform-agnostic logic. This architecture enables developers to swap platform-specific implementations through well-defined interfaces, leading to significant code reduction and maintenance ease without performance compromises. The approach has demonstrated substantial efficiency, as seen in operations like conv2d and block-scaled matmul, and is poised to extend further into areas like warp-specialized kernels and automated scheduling, thus forming a robust foundation for future high-performance computing across diverse hardware.

Google DeepMind has released the Gemma 4 family of models, which are state-of-the-art open multimodal models supporting text, images, and video, with enhanced performance available on both NVIDIA and AMD hardware through Modular Cloud. The Gemma 4 31B model boasts a 31-billion-parameter dense architecture with a 256K context window for complex tasks, while the Gemma 4 26B A4B is a Mixture-of-Experts model that activates only 4 billion parameters per pass to reduce compute costs. Modular Cloud offers a seamless transition from testing to production, leveraging the MAX inference framework to optimize performance and ensure consistency across different workloads. With 15% faster throughput on NVIDIA B200 compared to vLLM, Gemma 4 provides high efficiency without accuracy loss, making it one of the most capable open models available for developers and enterprises eager to deploy advanced AI solutions.