IBM announced the completion of its acquisition of data streaming platform Confluent for $31 per share in cash, with an enterprise value of approximately $11 billion. Based on Apache Kafka, Confluent is used by over 6,500 enterprises and serves about 40% of Fortune 500 companies. IBM plans to integrate it with watsonx.data, IBM MQ, webMethods hybrid integration, and IBM Z to deliver event-driven automation and real-time analytics across hybrid cloud environments, enabling enterprise AI and intelligent agents to trigger workflows and decisions based on real-time data rather than offline batch processing.

OpenAI has released two more hardware-efficient models in the GPT-5.4 series: GPT-5.4 mini and GPT-5.4 nano. The mini model reportedly achieves around 95% of GPT-5.4's performance on programming and computer operation tasks, supports a 400K-token context window and multimodal inputs, and is priced at $0.75 per million input tokens and $4.5 per million output tokens via API. The nano model targets lightweight tasks such as extraction and classification, available only via API, priced at $0.20 per million input tokens and $1.25 per million output tokens. Generating descriptions for approximately 76,000 images at this rate would cost about $52.

Rakuten Group has launched 'Rakuten AI 3.0,' one of Japan's largest high-performance language models, featuring a Mixture-of-Experts (MoE) architecture with approximately 700 billion parameters and optimized for Japanese. Developed as part of the GENIAC project led by Japan's Ministry of Economy, Trade and Industry (METI) and NEDO, the model was trained on Rakuten's proprietary high-quality bilingual data and excels in various Japanese-language benchmarks, including cultural and historical knowledge, graduate-level reasoning, competitive mathematics, and instruction following. The model is publicly released under the Apache 2.0 license and can be freely downloaded and used via its official Hugging Face repository.

Mistral AI has released and open-sourced Mistral Small 4, a model with approximately 119 billion total parameters using a MoE architecture—dynamically activating 4 out of 128 experts per token, involving around 6 billion active parameters. It offers both fast and deep-thinking modes, supports a 256K-token context window and image inputs, and includes configurable inference intensity parameters to balance cost and performance. Multiple reports indicate a 40% improvement in completion speed and a threefold increase in throughput compared to its predecessor. Released under the Apache 2.0 license, it is now available on Hugging Face, Mistral API, and NVIDIA platforms; self-hosting reportedly requires at least four H100-class GPUs.

Alibaba's DingTalk unveiled 'Wukong,' an AI-native enterprise work platform positioned as a unified agent entry point for business workflows, aiming to serve over 20 million organizations already on DingTalk. The company stated it has completely rewritten the underlying codebase, transitioning from a traditional GUI to a CLI and open API architecture, enabling AI agents to natively invoke DingTalk capabilities via commands. It emphasizes 'communication as execution' and transparent, human-intervenable task management. The platform also introduces RealDoc, a proprietary AI-native file system supporting precise file operations, change tracking, and instant version rollback, along with a six-layer enterprise security framework and initial OPT (One-Person Team) industry solutions, now entering invitation-only testing.

Citing OpenRouter platform data, 36Kr reported that during the week of March 9–15, 2026, Chinese large models were called for 4.69 trillion tokens, exceeding the U.S. volume of 3.294 trillion for the second consecutive week—up from 4.19 trillion vs. 3.63 trillion the previous week. The top three models were all Chinese: MiniMax M2.5 (1.75 trillion), Step 3.5 Flash (free) by Jieyue星辰 (1.34 trillion), and DeepSeek V3.2 (1.04 trillion). A new entrant, 'Hunter Alpha,' ranked seventh with 0.666 trillion token calls, reportedly featuring 1 trillion parameters and a 100K-token context, targeting agent applications.

Microsoft is reported to have restructured its Copilot organization, merging the Microsoft 365 Copilot and consumer Copilot teams, stating the move aligns product structure more closely with system architecture to deliver a more cohesive user experience. Jacob Andreou will lead Copilot product design, growth, and engineering across both consumer and commercial segments, while Mustafa Suleyman, CEO of Microsoft AI, will focus more on AI model development and the 'superintelligence' vision. The company also mentioned future plans to unify the Copilot experience across commercial and consumer versions to reduce fragmentation-related collaboration costs and improve product execution efficiency amid intensifying competition.

Kenya is advancing the '2026 Artificial Intelligence Act,' proposing the creation of an 'Office of the Artificial Intelligence Commissioner' as a national regulatory body, establishing a risk-based classification governance framework, and granting citizens rights to challenge automated decisions, request human review, and obtain explanations. The bill specifically targets misuse of generative AI to create misleading or harmful content, especially non-consensual deepfakes, with penalties of up to 5 million Kenyan shillings or two years' imprisonment—or both—for individuals or organizations. It also mandates clear labeling of AI-generated content to mitigate disinformation risks in political and public communication contexts.

Startup Niv-AI has emerged from stealth with $12 million in seed funding, aiming to improve GPU power efficiency. The company highlights that data centers often throttle GPU performance due to瞬时 power fluctuations, potentially wasting up to 30% of compute capacity. Its solution involves deploying rack-level sensors to monitor GPU power consumption with millisecond precision and modeling the power profiles of different training/inference workloads. Software tools then predict and coordinate overall power usage, helping engineers increase utilization without adding hardware and reducing strain on the grid. Niv-AI plans to deploy in a limited number of U.S. data centers within the next 6–8 months.

Researchers warn that attackers can embed 'semantic injection' instructions in open-source project READMEs or linked documents to trick AI agents into exfiltrating local sensitive data during installation or configuration steps. Testing on ReadSecBench—a dataset of 500 real-world open-source READMEs covering Java, Python, C, C++, and JavaScript—showed that direct command-style malicious prompts successfully triggered data leaks in about 85% of cases. When hidden two layers deep in external links, success rates remained as high as 91%. Tests included AI agents powered by Claude, GPT, and Gemini. The study notes that both human reviewers and existing automated detection tools struggle to identify such threats and recommends treating external documentation as 'partially trusted input' and introducing tiered verification for high-risk operations.