This past month brought a wave of community projects and milestones across the Modular ecosystem!

Modular Platform 25.5 landed with Large Scale Batch Inference, leaner packages, and new integrations that make scaling AI easier than ever. It’s already powering production deployments like SF Compute’s Large Scale Inference Batch API, cutting costs by up to 80% while supporting more than 15 leading models.

Around the world, the Modular community has been busy: experimenting with Gaussian splatting, building probabilistic data structures in Mojo, digging into GPU puzzles, and hosting meetups. Mojo even made its debut in the 2025 Stack Overflow Developer Survey, just two years after launch, another milestone in its rapid adoption.

Let’s take a look at everything the Modular universe made possible.

Blogs, Tutorials, and Videos

• At our July Community Meeting, Maxim presented his newly merged work on Hasher-based hashing, and we heard from all three Modular Hack Weekend winners: Martin Vuyk, who implemented Fast Fourier Transform in Mojo, Seth Stadick, who built Mojo-Lapper, a GPU-accelerated interval overlap detection library, and Thomas Trenty, who created QLabs, a GPU-powered quantum circuit simulator.

• The Mojo team shared a Q2 recap and Q3 roadmap update in the forum.
• Two Modular Platform applications are now live in AWS Marketplace under the new AI Agents and Tools category:
  • MAX High-Performance GenAI Serving Platform, with 500+ pre-optimized models, OpenAI-compatible APIs, and cross-platform GPU acceleration.
  • MAX Code Repo Agent, offering repository-aware Q&A, automated documentation, and intelligent code assistance.
• At the AI Performance Engineering Meetup, Ehsan M. Kermani showed how to replace any part of a PyTorch model with MAX and Mojo, making custom ops much easier.
• Shubham Gupta published a two-part deep dive into the GPU Puzzles series: Part 1 covers puzzles 1-8, and Part 2 explains puzzles 9-14. Once you’ve solved them all, let us know and we’ll send you stickers.
• New challenges are here for GPU devs:
  
  • ‍GPU Puzzle 25: use async memory operations to overcome GPU memory bottlenecks and optimize a memory-bound 1D convolution.
  • Part II of GPU Puzzles: puzzle 9 introduces debugging workflows and common GPU issues, while puzzle 10 teaches you to use NVIDIA’s compute-sanitizer to find and fix race conditions.
• Chris Lattner joined Richard Feldman on the Software Unscripted podcast.
• New Modular Tech Talk: Deep Dhillon explores the challenges of serving high-performance LLMs and how Mammoth, our Kubernetes-native distributed serving tool, delivers scalable performance across architectures.

• Mojo made its debut in the 2025 Stack Overflow Developer Survey, just two years after launch.
• We’ve partnered with SF Compute to launch the Large Scale Inference Batch API, offering up to 80% cost savings, support for 15+ leading models, and real-time GPU spot pricing. Watch the launch video, read the blog post, and get in touch if you’d like to try it.

• Modular Platform 25.5 is here, built for developers who need scale. Highlights include:
  • Large Scale Batch Inference, a high-throughput OpenAI-compatible API powered by Mammoth, already live in production with SF Compute.
  • Standalone Mojo Conda packages, leaner (<700 MB) MAX Serving packages, a fully open-source MAX Graph API, and seamless MAX ↔ PyTorch integration. Full details: MAX changelog | Mojo changelog.
• The August Community Meeting featured mojo-regex optimizations from Manuel, Apple GPU updates from Amir, and live Q&A with the team.

• Missed our Modular Platform 25.5 livestream? We covered Large Scale Batch Inference, Mojo, MAX, PyTorch custom ops, and fielded a ton of community questions.

• Big things are happening on August 28th at our Los Altos HQ: talks and networking with Modular and Inworld AI! Chris Lattner on the open future of compute and Mojo, Feifan Fan on voice AI in production, and Chris Hoge on matmul optimization. Grab your seat.

• Ferdinand Schenck wrote about translating scikit-learn’s Cython into Mojo.
• Ekemini Samuel launched a Mojo-focused community in Africa and hosted their first meetup in Uyo.

Awesome MAX + Mojo

• Bernardo shared his work porting Gaussian splatting kernels (written in CUDA C++) to Mojo.

• seiflotfy built mojo-hyperloglog, a Mojo implementation of HyperLogLog, which is a probabilistic data structure for counting unique elements with minimal memory usage.
• rd4com created live-stats-for-linux-based-operating-systems-in-mojo, an app that provides a live system overview of CPU and GPU stats.

• TilliFe released GPU support for the Nabla Python API.
• HammadHAB published a full API reference, installation documentation, and examples for his Mojo GUI library, CombustUI.

Open-Source Contributions

If you’ve recently had your first PR merged, message the Modular team in the forum to claim your epic Modular swag!

Check out the recently merged contributions from our amazing community members:

• mzaks [1][2][3]
• martinvuyk [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19]
• soraros [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35][36][37][38][39][40][41][42]
• ScottzCodez [1][2][3]
• zsiegel92 [1]
• LeeLee26 [1][2][3]
• Julian-J-S [1]
• Amila-Rukshan [1]
• msaelices [1][2]
• marsmxm [1]
• samufi [1]
• Caslyn [1]
• AceMouse [1][2][3]
• mmicu [1][2][3][4][5][6][7]
• benz0li [1]
• Amet13 [1]
• yeison [1]
• gustawdaniel [1]
• cyrillzadra [1][2]
• cnhz95 [1][2]
• christoph-schlumpf [1][2][3][4][5]
• Rtosshy [1]
• kyoto7250 [1]
• SasankYadati [1]
• simonyjung [1]
• Alex-Mann [1]
• cudawarped [1]
• farnoy [1]
• josiahls [1]
• rd4com [1]