Last Week in AIGoogleWednesday, February 4, 2026

Last Week in AI #334 - Kimi K2.5 & Code, Genie 3, OpenClaw & Moltbook

AI-Generated Summary

China's Moonshot has released Kimi K2.5, an open source model, alongside a new coding agent. This development represents continued competition in the large language model space, with Chinese companies advancing their capabilities in both general-purpose and specialized AI systems. The open source nature of Kimi K2.5 may facilitate broader adoption and development within the AI research community.

Google has launched Genie 3, an interactive world-building prototype, to its AI Ultra subscribers. This tool appears to represent a new frontier in generative AI applications, moving beyond text and code generation into creative environment design. The exclusive rollout to premium subscribers suggests Google is positioning advanced creative tools as a high-value service offering.

These developments highlight the accelerating pace of AI innovation across multiple domains and regions. Companies are expanding beyond foundational language models into specialized applications—coding agents, world-building tools—while also pursuing different distribution strategies, from open source releases to premium features. The simultaneous progress by both Chinese and American firms underscores the intensifying global competition in AI development.

Key Takeaways

China's Moonshot has released Kimi K2.
5, an open source model, alongside a new coding agent.
This development represents continued competition in the large language model space, with Chinese companies advancing their capabilities in both general-purpose and specialized AI systems.
The open source nature of Kimi K2.

Read the full article on Last Week in AI

Read on Last Week in AI

The Register23h ago

Growing void between enterprise and frontier AI puts open weights models in the spotlight

Google

A significant gap is emerging between the capabilities of frontier AI models—the largest and most advanced systems—and what most enterprise customers actually need for their business operations. Organizations increasingly recognize that the most expensive, cutting-edge models often provide unnecessary overhead, leading them to seek alternative solutions that offer better value propositions without sacrificing essential performance.

Simon Willison5h ago

Gemma 4 audio with MLX

Google

Google's Gemma 4 model is now available for audio processing on macOS through MLX, a machine learning framework optimized for Apple silicon. The 10.28 GB Gemma 4 E2B variant can be run locally using a command-line recipe that leverages MLX and the mlx-vlm library, enabling users to transcribe audio files directly on their machines without relying on cloud services.

TechCrunch1 day ago

Google and Intel deepen AI infrastructure partnership

Google

Google and Intel have announced an expanded partnership focused on co-developing custom chips designed specifically for AI infrastructure. The collaboration comes amid surging global demand for computing processors, which has created significant supply constraints across the technology sector.

Simon Willison1 day ago

Google AI Edge Gallery

Google

Google has released an official application called AI Edge Gallery that enables users to run Gemma 4 language models directly on iPhones without requiring internet connectivity. The app supports multiple model sizes, including the E2B and E4B variants of Gemma 4, as well as select models from the Gemma 3 family. The E2B model requires a 2.54GB download and reportedly delivers both speed and capable performance.

Google AI1 day ago

New ways to balance cost and reliability in the Gemini API

Google

Google has introduced new configuration options for the Gemini API that let developers optimize their spending versus performance guarantees. This matters because AI API costs represent a significant operational expense for builders, and the ability to tune reliability requirements means smaller companies can access powerful models affordably while enterprises can pay for guaranteed uptime when needed.