AI Digest

Alibaba Unveils Qwen3-Plus, Claiming Frontier Reasoning Performance
Alibaba released Qwen3-Plus, a large language model that reportedly matches or exceeds GPT-4o on several reasoning benchmarks, including MATH and GSM8K. The model is available under an open-source license, continuing Alibaba’s aggressive push to compete with Western frontier labs. Early benchmarks show particular strength in multi-step mathematical reasoning and code generation.

OpenAI Announces GPT-5 Preview for Enterprise Customers
OpenAI began rolling out a preview of GPT-5 to select enterprise partners, with a focus on improved long-context handling (up to 256K tokens) and reduced latency. The company claims the new model achieves a 40% cost reduction per token compared to GPT-4 Turbo. A broader public release is expected later this quarter.

EU Parliament Passes Landmark AI Liability Directive
The European Parliament voted to approve the AI Liability Directive, which establishes a strict liability framework for harm caused by high-risk AI systems. Companies will be required to prove their systems were not defective in the event of an incident. The directive is expected to take effect in member states by 2027.

Google DeepMind’s AlphaFold 3.0 Released to Researchers
DeepMind released AlphaFold 3.0, which now predicts protein interactions with small molecules, DNA, and RNA, not just protein folding. The update enables drug discovery researchers to simulate binding dynamics directly. Access is free for non-commercial use via a new cloud API.

Meta’s Llama 4 Training Details Leak, Showing Massive Scale
A leaked internal document suggests Meta trained Llama 4 on over 30 trillion tokens using a cluster of 100,000 H100 GPUs. The model reportedly achieves state-of-the-art performance on multilingual tasks, including low-resource languages. Meta has not commented on the leak.

Microsoft Launches Copilot for Healthcare with HIPAA Compliance
Microsoft announced Copilot for Healthcare, a suite of AI tools integrated into its Dynamics 365 and Azure for medical records and clinical decision support. The system is HIPAA-compliant and includes a real-time transcription and summarization feature for doctor-patient consultations. Several major hospital networks have begun pilot deployments.

China’s New AI Compute Rules Limit GPU Export to Cloud Providers
The Chinese government issued new regulations capping the number of high-performance GPUs that cloud providers can deploy for AI training, citing energy consumption concerns. The rules will affect major domestic providers like Alibaba Cloud and Tencent Cloud. Analysts expect a short-term bottleneck for large-scale training in China.

Runway’s Gen-4 Alpha Offers Real-Time Video Editing via Text Prompts
Runway launched Gen-4 Alpha, a video generation model that allows users to edit existing clips in real time using natural language commands. The model can change backgrounds, remove objects, or alter lighting without re-rendering the entire scene. Early user reviews praise the speed but note occasional artifacts in complex scenes.

Editor’s take: Today’s news underscores a dual trend: the rapid commoditisation of frontier AI models (via open-source releases from Alibaba and Meta) and the tightening regulatory and infrastructure constraints that will shape their deployment. The EU’s liability directive and China’s compute cap signal that the era of unconstrained AI scaling is giving way to a more structured, and contested, global governance landscape.