Kimi K3, Gemini 3.5 Delay, ARC-AGI 3 Breakthrough

Moonshot released Kimi K3, a 2.8 trillion parameter multimodal model with a 1 million token context window. The model is optimized for faster long-context decoding and agentic coding, and is currently available via Kimi's product and API, with open-source weights planned for release on July 27. This release shows that in the model scale race, Moonshot is trying to establish an advantage through parameter count and context length.

According to a report, Google delayed the release of Gemini 3.5 Pro due to needed improvements in performance, particularly in coding. This news caused Alphabet's stock to drop 4%. Google responded that the model is still in partner testing, along with an upgraded Flash model and other versions. This event highlights the central role of coding ability in the current AI model competition and the direct impact of performance optimization on product release timelines.

A new method called Schema helped frontier models achieve 99% accuracy on the ARC-AGI-3 public test set. ARC-AGI-3 provides an agent with a game environment but does not explain what it sees. Schema achieves this by having the model write the mechanics of each game as an executable program, then test against reality and plan within it. It controls how observations translate into a working model of the game, how predictions are tested against interaction history, and how plans are executed and revised. This result shows that structured approaches can significantly improve model performance on abstract reasoning tasks.

Model routing has become a hot topic because no one wants to pay frontier model prices for every token. The best routing is deeply task-specific: the more focused on a single workflow, the greater the gains in accuracy and cost. The challenge is determining which model can meet which task's requirements under cost and latency constraints. Meanwhile, Replit engineers tripled code output using an AI agent system without sacrificing quality. These agents handle tasks like reviewing pull requests, investigating incidents, and analyzing data, allowing engineers to focus on strategic initiatives.