OpenAI released GPT-5.4 on March 5, positioning it as a unified model for enterprise reasoning, programming, and agentic tasks, claiming a 33% reduction in single-statement error rate and an 18% drop in per-response error rate compared to GPT-5.2. The model is now available to ChatGPT Plus, Team, Pro users, and API developers, with a higher-performance GPT-5.4 Pro version also offered. OpenAI also launched a beta version of ChatGPT for Spreadsheets (an Excel and Google Sheets plugin) and integrated data services from FactSet, MSCI, Third Bridge, and Moody’s to enhance enterprise workflows.

OpenAI has expanded its standalone Codex programming application to the Windows platform, completing native support previously focused on macOS. The app runs on PowerShell and provides a Windows sandbox environment, supporting parallel multi-agent collaboration, background Automations, Work Trees for task isolation, and optional WSL integration. Skills modules allow scripts, instructions, and resources to be connected to specific toolchains. OpenAI states this approach enables safer execution of testing and repetitive processes within local development environments; users can log in via ChatGPT subscription or API key, with session history and configurations synced across systems.

Lin Junyang, technical lead of Alibaba's Tongyi Lab Qwen large model, announced his resignation on social media on March 3, followed by the departure of several core R&D members, drawing attention from the team and open-source community. Multiple reports attribute the trigger to internal restructuring and prolonged resource constraints, with employees raising concerns over compute allocation and hiring limitations. Alibaba CEO Eddie Wu held an emergency meeting and reaffirmed that Qwen remains the company’s top priority, stating organizational adjustments aim at expansion rather than contraction, and pledged increased investment. External investors worry team instability could delay critical initiatives by six months to a year.

YuanLab AI has released and claims to have open-sourced the multimodal MoE foundation model Yuan 3.0 Ultra, with 1 trillion total parameters and 68.8 billion activated parameters. Its LAEP layer employs adaptive expert pruning to dynamically remove low-utilization experts during pretraining, compressing scale from 1.5T to 1T, and combined with expert reordering achieves a 49% improvement in overall pretraining efficiency (up to 92.60 TFLOPS per GPU). During reinforcement learning, RIRM is introduced to suppress 'overthinking,' increasing training accuracy by 16.33% and reducing output token length by 14.38%. The model reports leading performance on enterprise benchmarks such as Docmatix 67.4% and ChatRAG 68.2%.

Kling announced the global full rollout of its Kling 3.0 series models starting March 5, featuring upgraded 'Motion Control 3.0.' The company claims significant improvements in character motion and facial consistency under complex camera conditions such as head turns, side profiles, occlusions, and multi-angle shots, resulting in more coherent video generation. Users can upload motion reference videos, first-frame images, and subject videos/images, combining them with prompts for multimodal control to increase controllability and determinism in generation. The model also emphasizes stable expressions and lip-sync during dynamic scenes like dancing and gymnastics, reducing facial distortion and drift issues.

The University of Tokyo's Matsuo-Iwasawa Research Lab, in collaboration with Sakura Internet, has released a Japanese medical-specific LLM, 'Weblab-MedLLM-Qwen-2.5-109B-Instruct,' and launched a conversational AI service. Based on Qwen-2.5-72B-Instruct, the model was further trained on Japanese medical literature and achieved 93.3% accuracy on the 2025 Japanese National Physician Examination benchmark, reaching approximately 98% when combined with RAG and majority voting. It scored an F1 of 85% on electronic medical record standardization tasks. The service is available for research use only from March 5 to August 31; the team plans to advance multi-institutional collaboration and safety evaluation mechanisms.

Netflix has announced the acquisition of InterPositive, an AI film production tools startup founded by Ben Affleck in 2022. The team will join Netflix, with Affleck serving as advisor. InterPositive advocates training custom models on raw footage already shot by film crews, providing post-production assistance rather than 'generating actors.' Applications include wire removal, shot reconstruction, filling missing frames, adjusting lighting, and enhancing backgrounds, emphasizing adherence to editing consistency and filmmaking rules, with built-in safeguards to protect creators’ intent and decision-making authority. Netflix did not disclose the transaction amount or specific terms.

Compute provider IREN disclosed orders for over 50,000 Nvidia B300 GPUs, aiming to expand its AI cluster to approximately 150,000 GPUs, increasing capacity by 50%. The new equipment is expected to come online in phases during the second half of 2026 at data centers in British Columbia, Canada, and Texas, USA. The company estimates that once fully deployed, it could support annualized AI cloud revenue exceeding $3.7 billion. To fund the expansion, IREN has raised about $9.3 billion through customer prepayments, convertible bonds, GPU leasing, and financing arrangements, and plans additional capital expenditures of around $3.5 billion. It also announced a follow-on offering of up to $6 billion, with shares down about 5% in pre-market trading.

Google has announced the establishment of Google AI Center Berlin as an offline hub for researchers, developers, and academic, industrial, and policy communities to collaborate and exchange ideas. The center will host events focused on AI agents, scientific computing, and healthcare, and establish long-term research partnerships with the Technical University of Munich (TUM) and Helmholtz Munich. Some media reports indicate the center is part of Google’s broader €5.5 billion investment plan in Germany, aiming to integrate cloud and data infrastructure to provide local startups and research institutions with easier access to collaboration spaces and resource connections.