Data Points: DeepSeek releases a hybrid reasoning model

Welcome back! In today’s edition of Data Points, you’ll learn more about how:

Alibaba’s new text and image editing features
Adobe’s tool to chat with your PDFs and other docs
Browsemaster, a new framework for agentic search
Surya, an IBM/NASA model that predicts solar weather

But first:

DeepSeek releases V3.1 with hybrid thinking modes

DeepSeek announced DeepSeek-V3.1, a 671 billion parameter MoE model that combines thinking and non-thinking modes through different chat templates. The updated model supports 128K token context length, with significant improvements in tool usage, agent tasks, and response efficiency compared to previous versions. DeepSeek-V3.1-Think achieves comparable quality to DeepSeek-R1-0528 while responding more quickly, achieving 93.7 percent on MMLU-Redux, 84.8 percent on MMLU-Pro, and a 2091 Codeforces rating. This is DeepSeek’s first model that can effectively handle both reasoning-intensive tasks requiring step-by-step thinking and rapid responses for simpler queries, potentially simplifying deployment for developers. The model is available on Hugging Face and ModelScope under an MIT license, and under metered pricing via DeepSeek’s API. (Hugging Face)

OpenAI launches low-cost ChatGPT Go plan in India

OpenAI introduced ChatGPT Go, a country-specific subscription plan exclusively for India, priced at Rs 399 (approximately $4.50) per month. The plan offers expanded GPT-5 access with better Indic language support, up to 10 times more messages than the free tier, daily image generation, file uploads, advanced data analysis tools, and custom GPTs. ChatGPT Go accepts UPI payments, making it more accessible to Indian users who previously needed debit or credit cards for subscriptions. India is now ChatGPT’s second-largest market, but provides comparatively little paid revenue. This mid-tier option bridges the gap between the free version and the more expensive ChatGPT Plus (Rs 1,999/month) and ChatGPT Pro (Rs 19,900/month) plans. This plan offers clues to OpenAI’s emerging markets strategy by offering localized pricing and features tailored to regional needs. (OpenAI)

Alibaba releases Qwen-Image-Edit for advanced image editing

Alibaba launched Qwen-Image-Edit, a 20 billion parameter model that extends the already released Qwen-Image to enable precise image editing with advanced capabilities text transformation. The model combines inputs from Qwen2.5-VL for semantic control and a VAE Encoder for appearance control, allowing both high-level semantic edits (like style transfer and object rotation) and low-level appearance modifications (such as adding or removing elements). The system offers bilingual text editing in Chinese and English, preserving original fonts and styles while making corrections or modifications, a capability that has historically been difficult for AI systems to master. The model is currently available through Qwen Chat’s Image Editing feature as well as through GitHub and Hugging Face under an Apache 2.0 license. (GitHub)

Adobe Acrobat launches PDF Spaces for AI document collaboration

Adobe introduced PDF Spaces, a new feature in Acrobat that transforms PDFs, Office 365 files, and web links into interactive knowledge hubs where users can use AI chat to extract insights and collaborate. Users can use the tool to organize scattered documents, generate summaries with citations, and create personalized AI assistants that can analyze content based on specific roles like analyst or instructor. Teams can share PDF Spaces including custom AI assistants, enabling colleagues to access the same knowledge base and AI-guided insights rather than just static files. Like Google’s NotebookLM, PDF Spaces puts conversational AI directly into document workflows, potentially changing how organizations manage and extract value from their data repositories. PDF Spaces is now accessible through Acrobat’s homepage and in a new application, Acrobat Studio. (Adobe)

Researchers develop BrowseMaster framework for complex web search tasks

BrowseMaster divides web search tasks between a planner agent that formulates strategies and an executor agent that retrieves information through programmatic code execution. The system achieved scores of 30.0 on BrowseComp-en and 46.5 on BrowseComp-zh benchmarks, outperforming several existing systems including OpenAI’s Deep Research on the Chinese benchmark. The executor can perform up to 244 tool calls in a single invocation using code-driven execution, compared to one call at a time for traditional agents. BCurrent AI agents often struggle with tasks requiring both broad information coverage and deep reasoning, achieving near-zero accuracy on challenging benchmarks; here, BrowseMaster offers a promising approach. (arXiv)

IBM and NASA unveil model to forecast solar storms

Surya analyzes high-resolution solar images to predict space weather events that can disrupt satellites, power grids, and GPS systems. The model, trained on nine years of data from NASA’s Solar Dynamics Observatory, achieved a 16 percent improvement in solar flare classification accuracy and can visually predict where flares will occur up to two hours in advance. Solar storms pose significant risks to modern infrastructure, with potential global economic losses of $2.4 trillion over five years according to Lloyd’s estimates, making accurate forecasting critical as society’s dependence on space-based technology grows. For AI engineers, Surya is unusual in the size of the input data and the architecture built to handle images of such complexity. The model is available on Hugging Face and GitHub under an Apache 2.0 license. (IBM)

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

Read this week’s issue of The Batch for in-depth analysis of news and research.

This week, Andrew Ng shared insights from a recent Buildathon hosted by AI Fund and DeepLearning.AI, where over 100 developers built functional AI-powered products in just a few hours, highlighting the fast-evolving landscape of agentic coding and rapid engineering.

“What excites me most isn’t just what can now be built in a few hours. Rather, it is that, if AI assistance lets us build basic but fully functional products this quickly, then imagine what can now be done in a week, or a month, or six months.”

Read Andrew’s full letter here.

Other top AI news and research stories we covered in depth:

China is reevaluating its stance on U.S. AI processors, asking Nvidia and AMD to prove that their high-end GPUs don’t pose a national security threat.
Alibaba’s Wan 2.2 video models introduced a new “Mixture of Video Experts” architecture that filters noisy inputs from clearer ones to improve video understanding.
OpenAI is partnering with Oracle to expand compute capacity, tapping into a $30 billion, 4.5 gigawatt data center initiative linked to the Stargate Project.
New research shows that larger models tend to memorize more bits from their training data, raising fresh questions about generalization versus memorization.