Moonshot lands a big one with Kimi K2.5: Qwen releases its own giant flagship reasoning model

In today’s edition of Data Points, you’ll learn more about:

• OpenAI’s move to win over scientists and scholars
• How U.S. workers are adopting (and not adopting) AI
• Moltbot, an open-source local automation agent
• Nvidia’s latest open weather and climate models

But first:

Kimi takes the open model crown, matching Sonnet at lower costs

Moonshot AI released Kimi K2.5, an open-weights multimodal model trained on approximately 15 trillion mixed visual and text tokens, building on its predecessor K2. The model can self-direct an agent swarm with up to 100 sub-agents that execute parallel workflows across up to 1,500 tool calls, reducing execution time by up to 4.5 times compared to single-agent setups. K2.5 is available through Kimi.com, the Kimi App, an API, and Kimi Code, with four modes including K2.5 Instant, K2.5 Thinking, K2.5 Agent, and K2.5 Agent Swarm in beta. Moonshot AI trained the agent swarm capability using Parallel-Agent Reinforcement Learning, which employs a trainable orchestrator to decompose tasks into parallelizable subtasks executed by dynamically created sub-agents. The model excels at coding tasks, particularly front-end development, and can generate complete interfaces from simple prompts or reconstruct websites from video input. (Moonshot AI)

New Qwen model matches the United States’ best

Alibaba’s Qwen team unveiled Qwen3-Max-Thinking, a flagship reasoning model that matches the performance of leading competitors like GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro across nineteen established benchmarks. The model incorporates adaptive tool-use capabilities that automatically select and invoke search, memory, and code interpreter functions during conversations without requiring manual setup from users. A novel test-time scaling strategy uses iterative self-reflection guided by a “take-experience” mechanism to boost reasoning performance more efficiently than standard parallel sampling, achieving improvements across multiple benchmarks including GPQA (90.3 to 92.8) and LiveCodeBench v6 (88.0 to 91.4). Qwen3-Max-Thinking is now available through Qwen Chat at chat.qwen.ai and via API for developers who register with Alibaba Cloud. (Qwen)

Prism, OpenAI’s scientific paper manager, is like Overleaf with AI

OpenAI released Prism, a free workspace that integrates GPT-5.2 directly into scientific writing and collaboration. The platform combines drafting, revision, and publication preparation in a single cloud-based environment built on LaTeX, eliminating the need for researchers to switch between separate editors, reference managers, and chat interfaces. Prism allows scientists to chat with GPT-5.2 to explore ideas, draft papers with full document context including equations and citations, search relevant literature from sources like arXiv, and convert whiteboard equations directly into LaTeX. The service is available now to anyone with a ChatGPT personal account and offers unlimited projects and collaborators at no cost, with plans to expand to ChatGPT Business, Enterprise, and Education users soon. (OpenAI)

Gallup poll shows sharp AI adoption curve, with caveats

American workers have rapidly adopted AI into their jobs, with twelve percent of employed adults using it daily at work, according to a Gallup survey of more than twenty-two thousand U.S. workers conducted this fall. The survey found that roughly one-quarter use AI at least a few times a week, and nearly half use it at least a few times a year, up sharply from 21 percent who were using AI at least occasionally in 2023. AI adoption remains concentrated in certain sectors, with about six in ten technology workers using AI frequently and three in ten doing so daily, while finance and education workers also report high usage rates. Workers across sectors use AI primarily for chatbots and virtual assistance, with four in ten also relying on it to consolidate information, generate ideas, or learn new things. However, economic researchers have identified roughly 6.1 million American workers who are heavily exposed to AI but poorly equipped to adapt, many of them in administrative roles concentrated in smaller cities. (Associated Press)

Clawdbot gets a new name and a higher profile

Clawdbot (now Moltbot) is an open-source personal AI agent that runs entirely on a user’s computer and integrates with multiple large language models like Claude Opus 4.5 and messaging services including Telegram, iMessage, and WhatsApp. The agent stores its settings, memories, and instructions as local folders and Markdown files on the user’s machine, giving it access to the filesystem and Terminal to execute commands, write scripts, and install new capabilities on demand. Users can ask Moltbot to improve itself with new features, such as adding voice transcription through the Whisper model or text-to-speech through ElevenLabs Eleven V3, and the agent will research documentation and implement the functionality autonomously. The system can control smart home devices like Philips Hue lights and Sonos speakers, manage productivity tools like Todoist and Notion, and replace cloud automation services by setting up cron jobs directly on the Mac. Because everything runs locally except for API calls to the chosen language model provider, users maintain full control over their data and can inspect or modify any aspect of the agent’s behavior. However, some security experts warn that giving Moltbot free rein can expose them to attacks. (MacStories)

Nvidia challenges Google in AI weather prediction

Nvidia unveiled the Earth-2 family of AI models and tools for weather and climate forecasting, marking what the company calls the world’s first fully open, accelerated weather AI software stack. The release includes three new models: Earth-2 Medium Range for 15-day global forecasts using a new Atlas architecture, Earth-2 Nowcasting for hyperlocal storm predictions in minutes using the StormScope architecture, and Earth-2 Global Data Assimilation for generating atmospheric snapshots in seconds instead of hours. The models outperform traditional physics-based supercomputer forecasting on standard benchmarks while drastically reducing computational costs. Earth-2 Medium Range and Nowcasting are available now via GitHub and Hugging Face, with Global Data Assimilation arriving later this year. (Nvidia)

Still want to know more about what matters in AI right now?

Read the latest issue of The Batch for in-depth analysis of news and research.

Last week, Andrew Ng talked about the need for businesses to redesign workflows to achieve transformative impacts with AI, rather than just incremental efficiency gains, and highlighted discussions at the World Economic Forum on topics like Agentic AI, Sovereign AI, and talent challenges.

“Against the backdrop of growing geopolitical uncertainty, I hope all of us in AI will keep building bridges that connect nations, sharing through open source, and building to benefit all nations and all people.”

Read Andrew’s letter here.

Other top AI news and research stories covered in depth:

• OpenAI tested advertisements for U.S. ChatGPT users in free and lower-cost tiers, marking a significant shift in its monetization strategy.
• Nvidia introduced Alpamayo-R1, a robotics-style reasoning model designed to enhance the decision-making capabilities of autonomous vehicles.
• Apple announced that its foundation model of choice would be Gemini, partnering with Google to enhance Siri and other AI-driven features.
• FlashWorld generated 3D objects, scenes, and surfaces with photorealistic fidelity, advancing the capabilities of text- or image-to-3D conversion.

A special offer for our community

DeepLearning.AI recently launched the first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:

• Over 150 AI courses and specializations from Andrew Ng and industry experts
• Labs and quizzes to test your knowledge 
• Projects to share with employers 
• Certificates to testify to your new skills
• A community to help you advance at the speed of AI

Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one week free trial. Explore Pro’s benefits and start building today!

Try Pro Membership