Modular publishes the complete MAX source under a new permissive license
AspireAI · Modular
Modular, the company behind the Mojo language, announced on October 22, 2026 that the complete MAX source code is now publicly available: the graph compiler, the runtime and the hardware backends, joining the kernels, models and serving infrastructure the company opened earlier. NVIDIA, AMD, Apple Silicon and CPU backends ship as reference implementations for anyone bringing MAX up on a new accelerator, and Modular plans to release plugins for Qualcomm Dragonfly and Google TPU in coming months. Permissions are broad - build on any hardware, no licensing fees or royalties, ship commercial products - and many components stay Apache 2.0, including the Mojo kernel library, accelerator interfaces, the serving and modeling Python APIs, and the models themselves. The restrictions are what make it 'source available' rather than open source: under the new MAX License you cannot use any part of MAX as input to an AI system to produce a reimplementation, cannot publish or distribute modified MAX source except by forking for a contribution back to the project, and must prominently display a 'Modular Accelerated' wordmark if you sell hosted inference or training services. Modular's stated rationale is anti-fragmentation - it says it hopes to replicate LLVM's success of getting people to build on a shared codebase. Whether that protects a shared foundation or a moat is a fair question. This completes a sequence: Mojo's standard library in 2024, MAX kernels and serving through 2025, and the full Mojo compiler in August 2026.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.