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

· Source: Last Week in AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Software Development & Engineering · Depth: Intermediate, medium

Summary

Moonshot AI has launched Kimi K2.5, an open-source, natively multimodal model trained on 15 trillion mixed visual and text tokens, capable of understanding text, images, and video. K2.5 demonstrates strong agentic capabilities, outperforming Gemini 3 Pro on SWE-Bench Verified, and both GPT 5.2 and Gemini 3 Pro on SWE-Bench Multilingual. For video understanding, it surpasses GPT 5.2 and Claude Opus 4.5 on VideoMMMU. Additionally, Moonshot introduced Kimi Code, an open-source coding agent that translates UI designs from images or videos into code, supporting integration with editors like VSCode. Google is expanding access to Genie 3, an experimental "general-purpose world model," to AI Ultra subscribers, enabling generation of dynamic, navigable 3D worlds from text and images. Meanwhile, OpenClaw (formerly Moltbot), an open-source, always-on AI assistant, has gained significant traction for its multi-platform messaging integration, despite security concerns regarding its access to real-world applications.

Key takeaway

For Machine Learning Engineers evaluating new model architectures, consider Moonshot AI's Kimi K2.5 for its multimodal capabilities and strong benchmark performance in coding and video understanding. Your teams should investigate its potential for agentic applications and UI-to-code generation, especially if seeking open-source alternatives to established models. Be mindful of the security implications when deploying always-on AI assistants like OpenClaw that access real-world applications.

Key insights

Multimodal AI models and coding agents are rapidly advancing, alongside new interactive world-building and always-on AI assistants.

Principles

Method

Reinforcement Learning via Self-Distillation uses the model as an on-policy "self-teacher" by conditioning on tokenized feedback to produce dense, logit-level supervision for policy updates.

In practice

Topics

Best for: Machine Learning Engineer, Computer Vision Engineer, CTO, AI Engineer, AI Product Manager, AI Researcher

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Last Week in AI.