AI Daily | AlphaProof Conquers Math, Grok V9, MiniCPM5-1B, and NuExtract3 Analysis

New Breakthroughs in AI: AlphaProof Solves Math Riddles and Grok V9 Enhances Coding Abilities

This article provides a detailed look at recent major advancements in the field of Artificial Intelligence. It covers DeepMind’s success in solving a half-century-old mathematical problem, as well as the latest technical and practical applications of the Grok V9, MiniCPM5, and NuExtract3 models, offering readers a glimpse into how these technologies are changing the future of computing.

To be honest, watching the progress of AI can be breathtaking. New computing models and algorithms are emerging like mushrooms after rain. From major breakthroughs in theoretical mathematics to the constant iteration of edge-device application models, the overlapping development of various technologies is dazzling. Here, we’ll detail several representative AI developments and explore the substantive changes these technologies bring.

A Shockwave in Mathematics: AlphaProof Nexus Conquers Half-Century-Old Challenges

Unsolved mysteries that have been sealed for decades are now being solved one by one by algorithms. This sounds like the plot of a science fiction novel, but it’s actually happening.

According to the paper Advancing Mathematics Research with AI-Driven Formal Proof Search, the AlphaProof Nexus system developed by Google DeepMind has successfully and autonomously solved 9 open Erdős math problems. Two of these problems had been unsolved for 56 years. You might wonder, what does this represent? While previous language models were smart, they often suffered from logical hallucinations when dealing with rigorous mathematical proofs. This new system ingeniously combines Large Language Models with the Lean formal language, allowing a compiler to automatically verify every logical step, ensuring the absolute correctness of the proof.

When mentioning mathematical proofs, most people might imagine a blackboard full of complex formulas, which can be somewhat intimidating. However, this is exactly where a logically rigorous language model can shine. AlphaProof Nexus uses a very special architectural design. The system contains multiple sub-agents that operate independently to find proofs. More advanced versions even introduce evolutionary algorithms, allowing the model to learn from past attempts and continuously evolve.

What’s surprising is the cost-effectiveness of the entire process. For these extremely difficult math problems, the inference cost to solve a single problem is only a few hundred dollars. Additionally, the system successfully proved 44 out of 492 conjectures in the On-Line Encyclopedia of Integer Sequences (OEIS). This undoubtedly brings a new auxiliary tool to mathematical research, allowing researchers to focus more on conceptual ideation.

Grok V9-Medium Training Completed: The Next Step in Strengthening Coding Capabilities

Beyond the impact on academic research, the industry is moving just as fast. Elon Musk recently posted about the completion of Grok V9-Medium training, sparking intense discussion in the tech community.

This base model, V9-Medium with 1.5 trillion parameters, has completed its preliminary training, and current evaluation data looks excellent. A massive amount of data from Cursor was added during the supplementary training phase. Readers familiar with development tools know that Cursor has a very high reputation in the field of code-assisted editing. This move is clearly intended to significantly boost Grok’s ability to handle complex coding tasks.

Fine-tuning is currently in full swing, and the reinforcement learning phase will begin within a few days. It’s expected that in two to three weeks, this model will be officially released to the public. Compared to the current 0.5 trillion parameter v8-small version that handles all of Grok’s production traffic, V9-Medium will bring a huge performance leap. Especially for logically tedious coding tasks that require high contextual understanding, the new version is expected to demonstrate much stronger support.

The Edge Inference Giant: MiniCPM5-1B Debuts

While mentioning the progress of large models, we absolutely cannot ignore those small models that perform brilliantly in resource-constrained environments. After all, many practical application scenarios don’t have unlimited cloud computing resources to spare.

Launched by OpenBMB, this 1-billion parameter model is designed specifically for terminal devices and local deployment. You can check the MiniCPM5-1B project page for detailed information. This dense Transformer model reaches top-tier standards among open-source models of the same scale. It is particularly good at using agent tools, code generation, and difficult logical reasoning.

The model introduces a Hybrid Reasoning mechanism with built-in thought-mode chat templates. Users can freely switch the model between a fast-responding assistant and a deliberate reasoner based on their needs. The development team adopted a fine-grained data-level management strategy for training, combining supervised fine-tuning and reinforcement learning. For developers who want to run intelligent applications locally, you can refer to its GitHub resources for deployment, or go directly to the online demo platform to test its actual performance.

The Synergy of Structured Data and OCR: NuExtract3 Vision-Language Model

In everyday development and enterprise applications, handling complex documents is often the biggest headache. From PDF files, screenshots, and forms to receipts, how to accurately capture information has always been a challenge. Here’s another very practical new tool.

According to the NuExtract3 release announcement, the NuMind team has launched a 4-billion parameter vision-language model based on Qwen3.5-4B. It uses the Apache-2.0 license and its biggest feature is the perfect combination of structured data extraction (outputting JSON) and content extraction (OCR functionality outputting Markdown) in a single model.

If you have used the practical tool NuMarkdown, then NuExtract3 is its comprehensive upgrade. The development team has given this model excellent extraction reasoning capabilities through reinforcement learning, and this reasoning function can be turned on or off at any time according to task requirements.

To give the model excellent long-text understanding, the development team used 8 H100 GPUs for 3 days of training. The hardware requirements for this model are very accessible, needing only about 4GB of VRAM to run smoothly. At the same time, the official team provides various weight quantization formats such as Safetensors and GGUF. Readers can go directly to the free Hugging Face Space to try it out without registration. For further integration, you can also check the Hugging Face model page and related model collections for more deployment details.

Frequently Asked Questions (FAQ)

To help readers better grasp the main points of this article, here are several common questions and answers.

Q1: What is the significance of AlphaProof Nexus solving Erdős math problems? This achievement proves that large language models combined with formal verification tools can indeed avoid logical hallucinations. The system solved math puzzles that had been unsolved for over half a century at extremely low inference costs, providing a valuable automated auxiliary tool for future mathematical theoretical research.

Q2: When is Grok V9-Medium expected to be officially released? The model has completed base training and included Cursor data, and is currently undergoing reinforcement learning and fine-tuning. It is expected to be released to the public within two to three weeks, significantly improving the processing power for complex coding tasks.

Q3: In what scenarios is MiniCPM5-1B suitable? This 1-billion parameter model is designed for resource-constrained terminal devices and local deployment. It features hybrid reasoning, making it very suitable for developing local code assistants, lightweight agent tools, and edge computing scenarios that require logical reasoning.

Q4: How is NuExtract3 different from traditional OCR tools? NuExtract3 is a vision-language model that features both structured extraction and content extraction. It can not only convert document images into Markdown format but also extract precise JSON data according to specified templates, excelling particularly in handling documents with tables, forms, and complex layouts.