MiniMax has officially open-sourced its M2.7 model, achieving industry-leading performance across multiple benchmarks, scoring 56.22% on the SWE-Pro benchmark and 57.0% on Terminal Bench 2. M2.7 is MiniMax's first model deeply involved in its own evolution process, capable of building complex agent frameworks and completing highly sophisticated productivity tasks through agent teams, advanced skills, and dynamic tool search. The model is compatible from day one with chip platforms including Huawei Ascend and NVIDIA, and users can access it via Hugging Face. This marks a significant advancement for domestic large models in the open-source ecosystem and agent capabilities.

DeepSeek has launched a new version featuring a dual-track design with 'Fast Mode' and 'Expert Mode'. Fast Mode is optimized for everyday Q&A with emphasis on response efficiency, while Expert Mode focuses on complex reasoning and domain-specific analysis in finance, law, and scientific research, enhancing answer authority and explainability through dynamic knowledge retrieval, multi-step reasoning, and citation tracing. This move represents a key step in DeepSeek’s transition from a general-purpose conversational tool to deep vertical services, aiming to differentiate itself from products like ChatGPT Plus and Claude Pro by establishing a 'technical depth' brand identity.

Led by SoftBank, eight major Japanese companies—NEC, Honda, Sony, three major banks, and Nippon Steel—have formed a new company, 'Japan AI Foundation Model Development,' aiming to build a domestically developed trillion-parameter-scale AI foundation model by 2030. The model will support multimodal processing of text, images, audio, and video, and target 'Physical AI' applications such as autonomous robot control. The Japanese government, through the Ministry of Economy, Trade and Industry, plans to invest approximately 1 trillion yen over the next five years, with the new company applying for support under NEDO programs. This initiative aims to counter U.S. and Chinese leadership in AI and promote Japan’s AI self-reliance.

Sri Viswanath, former partner at Coatue, has founded Sycamore, an enterprise AI agent startup, securing $65 million in seed funding led by Coatue and Lightspeed. Viswanath previously held roles at Sun Microsystems, VMware, and Atlassian. Sycamore builds an AI agent 'orchestration layer' that fully manages coding and backend infrastructure tasks, distinguishing itself from conventional approaches that layer AI agents atop existing workflows. Angel investors include Bob McGrew, former Chief Scientist at OpenAI, and Lip-Bu Tan, CEO of Intel, signaling strong industry recognition of its technical direction.

Lingchu Intelligence has released Psi-R2, an embodied AI framework that leads the MolmoSpaces leaderboard using nearly 100,000 hours of human operation data and a dual-model architecture. Psi-R2 features a collaborative design of a policy model and a world model: the policy model learns successful task trajectories, while the world model Psi-W0 simulates failure scenarios via 'dream trial-and-error'. Training data spans 294 scenarios and 4,821 tasks, including 5,417 hours of real-robot data. The company improved pose accuracy to sub-millimeter levels using a proprietary exoskeleton glove and has open-sourced a 1,000-hour high-quality human manipulation dataset to advance the embodied AI ecosystem.

Security researchers have disclosed three critical vulnerabilities in LangChain and LangGraph: CVE-2026-34070 enables arbitrary file access via path traversal; CVE-2025-68664 allows API key and environment variable leakage through deserialization; and CVE-2025-67644 is an SQL injection vulnerability in LangGraph’s SQLite checkpoint system. Given the widespread use of LangChain and LangGraph in AI development frameworks, these flaws could trigger cascading impacts across the AI tech stack. A prior vulnerability in Langflow was exploited within 20 hours of public disclosure, underscoring the urgency of rapid patching.

Beijing Zhongke Wenge Technology Co., Ltd. officially filed its IPO prospectus on April 12, planning to list in Hong Kong with CICC as sole sponsor. Originating from the Institute of Automation, Chinese Academy of Sciences, the company specializes in enterprise AI and decision intelligence services, offering large models such as Yadi and Panstone. Shareholders include state-backed institutions like China Development Finance, CITIC Capital, and Shenzhen Capital Group. Revenue grew from 250 million to 405 million RMB between 2023 and 2025, with gross margins exceeding 51%, though cumulative losses exceeded 500 million RMB over the same period. Its R&D team comprises 250 members, holding 154 patents and 439 software copyrights. Proceeds will fund R&D, product expansion, and overseas growth.

French AI startup Mistral has secured $830 million in debt financing to build a data center near Paris, advancing European AI infrastructure and ensuring data sovereignty. The company aims to deploy 200 megawatts of compute capacity in Europe by 2027, using NVIDIA GB300 GPUs to create customized AI environments and reduce reliance on third-party cloud providers. This move is seen as a strategic effort to compete with U.S. AI giants and close the investment gap between Europe and America, marking progress in Europe’s push for autonomous AI ecosystems.

A research team has introduced ASI-Evolve, an agent framework that automates foundational AI research through a 'learn-design-experiment-analyze' loop. By integrating a persistent cognitive library of human priors with dedicated analyzers, the system achieves autonomous discovery in model architecture, pre-training data selection, and reinforcement learning algorithms. It successfully identified 105 novel linear attention architectures outperforming baselines, optimized data filtering strategies that boosted MMLU performance by over 18 points, and proposed new RL update rules surpassing existing baselines. This work marks a critical step toward recursive self-improvement in AI.

The European Commission is reviewing whether OpenAI's ChatGPT should be classified as a large online search engine, subjecting it to stricter regulation under the Digital Services Act (DSA). OpenAI disclosed that ChatGPT has approximately 120.4 million monthly active users in the EU, far exceeding the 45 million threshold. A Commission spokesperson stated that large language models require case-by-case assessment and will not be automatically categorized. If formally designated, ChatGPT would face enhanced transparency and risk management obligations. Meanwhile, certification challenges for agent-based AI under the EU AI Act are also drawing attention, with compliance deadlines for high-risk systems set for August 2026.