Google has upgraded its reasoning model Gemini3 DeepThink, enhancing 'deep thinking' capabilities for complex and information-incomplete problems. The model is now available to AI Ultra subscribers, with early API access granted to select researchers, engineers, and enterprises. Public benchmarks include HLE 48.4%, ARC-AGI-2 at 84.6%, Codeforces programming Elo of 3455, and gold medal-level performance in written exams for math, physics, and chemistry Olympiads. Demonstrations include capabilities such as generating 3D printable files from sketches, applicable to scientific peer review and experimental workflow optimization.

Mistral has launched Forge, an enterprise model development platform designed for large organizations requiring compliance and data sovereignty. It supports pre-training, post-training, and reinforcement learning on proprietary documents, structured data, code repositories, and operational records. The platform supports both dense and Mixture-of-Experts (MoE) architectures and multimodal inputs, emphasizing that enterprises retain ownership of models and data. It includes built-in evaluation and reinforcement learning pipelines for continuous alignment with internal standards. Forge is currently available to select partners, with disclosed users including ASML, Ericsson, and the European Space Agency.

Google has extended access to Gemini's Personal Intelligence feature for all free users in the U.S., who must opt in to enable the assistant to leverage account data from Gmail, Drive, Maps, and other services to generate personalized recommendations—such as troubleshooting device issues based on purchase history or planning routes using travel and hotel bookings. Powered by the Gemini3 series, the system features a context window of up to 1 million tokens, using 'context packaging' to filter key information and adapt to long-context limitations. This capability was previously limited to Pro ($19.99/month) and Ultra ($249.99/month) subscription plans.

According to Reuters, citing sources, Nvidia has received approval from Beijing to resume selling H200 AI chips to China; the U.S. had previously allowed limited exports of H200 in February. Jensen Huang stated that the company has received orders from multiple Chinese customers and restarted supply chain production. The report also notes that Nvidia is preparing a variant of the Groq inference chip tailored for the Chinese market, expected to launch in May, targeting inference workloads and compatible with other systems. The Chinese market previously accounted for approximately 13% of Nvidia's total revenue.

Baidu has released the 'Lobster' family of AI agents for the OpenClaw open-source framework, positioning them as assistants capable of executing multi-step tasks across applications and devices, supporting desktop, cloud, mobile, and smart home environments. Use cases include video editing, presentation creation, information retrieval, and placing orders. Following strategic moves by Alibaba and Tencent into the OpenClaw ecosystem, Baidu aims to use agents as a new growth driver, emphasizing their potential to become a 'next-generation OS' that bridges hardware and software silos.

Tencent stated it has completed intensive upgrades to its AI team, organizational structure, and workflows over recent months, rebuilding its pre-training and reinforcement learning infrastructure to improve data quality and model iteration efficiency. President Liu Zhiping revealed that the Hunyuan large model 'Hunyuan 3.0' is currently in internal testing, with a phased public rollout planned to begin in early April. The company emphasized ongoing investment in multimodal capabilities, particularly in video and world model research. Its AI assistant 'Yuanbao' met or exceeded installation targets during the Lunar New Year campaign, and future efforts will focus on improving retention and engagement through experience optimization and ecosystem integration.

MiniMax has released its self-developed model M2.7, introducing a 'self-evolution' pathway: the model can automatically collect feedback via AgentHarness, build evaluation datasets, and iterate within an 'analyze-plan-modify-evaluate' closed loop, reportedly covering 30%–50% of the reinforcement learning development pipeline. The company also highlights SRE-grade engineering capabilities, enabling the model to locate issues by integrating monitoring and database logs, generate repair scripts, and reduce fault recovery time to under three minutes. M2.7 is designed for production integration, emphasizing multi-agent collaboration and compatibility with developer toolchains.

S&P Global announced the acquisition of Enertel AI to strengthen its power market data and decision-making tools. Enertel specializes in short-term, node-level electricity price forecasting for North American markets, covering day-ahead and sub-hourly intervals, using graph neural networks to integrate variables such as market prices, load, weather, and fuel costs. S&P Global plans to combine this real-time forecasting capability with its long-term energy intelligence, offering power traders, utilities, and asset operators an integrated view from long-term strategy to next-day node pricing. The company stated the transaction is not expected to have a significant impact on overall financial performance.

Singapore-based generative video startup Video Rebirth announced total funding of $80 million, including a $30 million extension round in March (following a $50 million raise in November 2025). The round included strategic investors such as AMD Ventures, Hyundai Motor Group, and CJ Group. Funds will support commercialization of the Bach series of video generation products and international expansion. The company claims its 'industrial-grade AI engine' uses Dual DiT and Physics-Native Attention techniques, supporting native 30fps generation to enhance consistency and controllability in long-form videos, targeting applications in entertainment content, simulation, and mobility.