Agent & Model Releases
Meituan released LongCat-2.0, a 1.6T-parameter open MoE model designed for agentic coding, featuring a native 1M context window and trained on over 35 trillion tokens using domestic AI ASIC superpods. The model activates only 48B parameters per token, supporting code understanding, generation, and execution in agent workflows.
Tools & Frameworks
A new tutorial walks through end-to-end GRPO training on Gemma-3 for structured math reasoning using Tunix, JAX, and LoRA adapters. The setup keeps training lightweight with adapter-only updates and custom reward functions, enabling reinforcement learning-based improvement on a single accelerator.
Industry & Business
Amazon will stop accepting new customers for Mechanical Turk on July 30, 2026, effectively winding down the crowdsourcing platform. Existing customers can still use the service, but no new features are planned, marking an end to a tool once crucial for AI data labeling and annotation.
Research Highlights
Researchers introduced Asymmetric Mutual Variational Learning (AMVL), a framework that resolves train-inference mismatch in multimodal continuous reasoning. By using a bidirectional calibration objective, it prevents answer leakage and improves reasoning without losing perceptual detail.
That's everything for this day – about a 5-minute read.