Research presented at ICML 2026 introduced selective activation sparsity, a training method enabling models to use only the most relevant parameters for specific tasks. Models trained with this technique matched the performance of models three times their size on reasoning benchmarks, with significant implications for training and inference costs on resource-constrained devices like phones and laptops.

Apple released new Intelligence frameworks on July 28 supporting on-device AI capabilities with image inputs, server model integration, and custom skill building. The Foundation Models framework now provides a single native Swift API enabling developers to integrate powerful models from Apple and third parties directly into applications, democratizing on-device AI.

Moonshot AI released Kimi K3 with 2.8 trillion parameters, the largest openly available LLM to date, published on July 26. Thinking Machines shipped its first-ever model, Inkling (975B parameters), on the same week. These releases represent a significant escalation in open-source capability and mark a tightening US-China model contest, with open weights now matching closed-model performance in specific domains.

The physical AI sector accelerated dramatically with Figure AI's Figure 03 crossing 1,000 units at BotQ (July 23) at 1 robot/hour throughput, while AgiBot reached 15,000 cumulative units. Boston Dynamics' electric Atlas began deployments, and multiple companies reported successful real-world factory deployments. The industry momentum shifted decisively toward logistics and manufacturing applications rather than experimental prototypes.

AMD released the Ryzen AI Embedded X100 series and Kria AI SoM (System on Module) with unified memory architecture and real-time deterministic control, positioned to outperform NVIDIA's Orin and Thor in robotics workloads. The move represents AMD's full-stack play in AI hardware and marks intensifying competition in the robotics accelerator market.