Kimi K2.5’s Agent Swarm & DeepSeek-OCR-2 Redefine AI Capabilities as ChatGPT Containers Gain Bash Power

New AI Models

Kimi K2.5 Released: Kimi K2.5, an open-source AI model, achieves state-of-the-art performance in agentic tasks, coding, image processing, and video analysis. It introduces an "agent swarm" feature for parallel processing and is available on the Kimi platform for chat, agent modes, and API access.

• Kimi K2.5 Released!!!
  • Blog Post
• Introducing Kimi K2.5, Open-Source Visual Agentic Intelligence
  • Website
  • Code
  • Platform
  • Blog Post
  • Hugging Face

DeepSeek-OCR-2 Released: DeepSeek-OCR-2, a new OCR model, offers advanced optical character recognition capabilities and is available on Hugging Face with a demo website.

• deepseek-ai/DeepSeek-OCR-2 · Hugging Face
  • Hugging Face

Jan-v3-4B-Instruct Released: Jan has launched Jan-v3-4B-base-instruct, a 4B-parameter AI model optimized for coding and general tasks, with a 40% improvement in Aider performance. It is available via Jan Desktop and Hugging Face.

• Jan v3 Instruct: a 4B coding Model with +40% Aider Improvement
  • Jan Website
  • Hugging Face (Base)
  • Hugging Face (GGUF)

AI Product Updates

ChatGPT Containers Enhanced: ChatGPT Containers now support running bash commands, installing packages via pip/npm, and downloading files, enabling more complex and interactive coding environments.

• ChatGPT Containers can now run bash, pip/npm install packages, and download files
  • Blog Post

AI Research & Experiments

Autonomous Agent Self-Refactor: Qoder's autonomous agent, Quest, successfully refactored its own codebase in 26 hours with minimal human intervention, rebuilding its interaction layer, state management, and core agent loop. The team is hosting an AMA to discuss the experiment.

• Our Agent Rebuilt Itself in 26 Hours. AMA👀
  • AMA Link