November 2025 Summaries

Modular Platform 25.7 introduces significant updates aimed at enhancing the performance and accessibility of AI compute layers, featuring a fully open MAX Python API and a new experimental modeling API that simplifies the development of high-performance inference models. The update also broadens hardware support, including NVIDIA Grace superchips, and introduces safer GPU programming through the Mojo language, which now features improved error detection and expanded Apple Silicon GPU support. Dynamic LoRA support is also introduced for real-time model specialization, particularly beneficial for speech and low-latency applications. These advancements position MAX as a leading inference engine, offering improved throughput and performance with a focus on openness and developer involvement. The release encourages community participation and feedback, emphasizing its commitment to building a unified and portable AI platform.

Inworld's collaboration with Modular has resulted in the Inworld TTS 1 Max model achieving the top position on the Artificial Analysis speech leaderboard, a platform that benchmarks over 100 large language models (LLMs) based on intelligence, speed, and cost. This ranking is determined by user preferences in real-world categories such as Customer Service and Entertainment, with users choosing between outputs without knowing their origins. The Inworld TTS 1 and TTS 1 Max models, supporting 12 languages and featuring advanced capabilities like voice cloning and emotion tags, are powered by the Modular platform using transformer-based, autoregressive models. The partnership has led to significant performance improvements, including a 70% faster API response for synthesized audio and a 60% reduction in costs, showcasing a promising future for accessible AI infrastructure.

Recent developer events, the PyTorch Conference and the LLVM Developers' Meeting, highlighted the convergence of challenges faced by the AI and compiler communities, particularly in grappling with hardware diversity and the need for integrated solutions across various layers of software development. At the PyTorch Conference, discussions focused on the hurdles developers face with multiple languages and frameworks, emphasizing the role of large language models (LLMs) in simplifying kernel development. Meanwhile, the LLVM Developers' Meeting showcased the adoption of MLIR to address hardware fragmentation and the need for long-term solutions in compiler technology. Both events underscored a shared desire for tools that balance performance, portability, and productivity amid rapid advancements in AI models and hardware. The Modular Platform, with its Mojo language and MAX inference framework, was presented as a potential solution, aiming to unify developers across different backgrounds and skill levels while accommodating diverse hardware needs.