Blog

In Part 1, we argued that LLM routing is qualitatively different from HTTP routing. Inference backends hold state that traditional load balancers ignore. This post covers the first of the three layers we identified: the data layer that makes that state queryable on the hot path of every inference request.

To build it, I needed libraries that did not exist yet or did not support the exact required features. So I built them:

Every millisecond matters in real-time voice, and at Hippocratic AI's scale latency gains compound directly into better patient experience and per-node efficiency. Production deployments run across multiple frameworks, including SGLang and vLLM, with ongoing evaluation of emerging frameworks for additional latency headroom, alongside a hardware roadmap spanning NVIDIA, AMD, and future-generation accelerators.

At Modular, we’re always experimenting with the latest agentic programming tools, integrating the best ones into our workflows, and learning quite a few lessons along the way. One thing we realized is that the Mojo language is ideally suited to the needs of modern AI coding agents.

Inkwell is a web app that lets users create interactive storybooks with a custom character along infinite branching paths. When the user opens a story, the first page of text and image art streams in - text appears character-by-character via WebSocket within the first second, the illustration paints in as you read, and by the time you tap a choice, the next page is already written and illustrated. Creating a user experience around the seamless generation of new content requires an inference layer that can perform at scale.

HTTP routing has been a solved problem for many years. Round-robin, consistent hashing, least-connections. Pick one, put it in front of a pool of identical servers, and the traffic spreads pretty evenly.

Surprise: Mojo 1.0 is officially in beta! Modular’s 26.3 release includes new features and modalities, but the headline is that we’ve officially hit beta for Mojo 1.0, with a clear plan to finalize Mojo 1.0 in the coming months. We share details below, alongside other key announcements in our 26.3 release including video generation in MAX with Wan 2.2 and MAX framework updates.

There was a lot to celebrate in April: the community shipped GPU renderers, FFmpeg bindings, raylib wrappers, BLAS routines, and a 2D graphics API, just to name a few. The team connected with tons of developers at AMD AI DevDay and our joint meetup with AMD, two new Modular offices opened on two different continents, and Gemma 4 launched with same-day support on NVIDIA and AMD. Here’s the April roundup.

In a clear demonstration of how rapidly AI coding agents are becoming capable of challenging systems engineering work, two of the five agents produced a working MAX pipeline. The models we tested were:

Suppose you want to load a 2D tile of a matrix, where the tile is stored in shared memory in a specific interleaved layout to avoid bank conflicts. This example uses a toy XOR swizzle to illustrate the class of bugs; real kernels use hardware- and layout-specific swizzles and vectorized accesses. Without a layout abstraction, here is how you would launch a kernel with a block size of (32,8):