Marvell announced plans to acquire optical communications company Lumentum in an all-stock deal valued at approximately $115 billion, combining digital signal processing, custom chips, lasers, and optical transceiver businesses to expand in the AI data center interconnect market. The companies expect annual cost synergies of more than $1 billion after the merger. Lumentum also reported fiscal first-quarter 2027 revenue of $1.263 billion, up 136.5% year over year, and secured an optical interconnect project order accompanied by a $5 billion purchase commitment.

Meituan has released LongCat-2.5, a model with approximately 1.6 trillion total parameters and around 480 billion parameters activated per token. It combines image recognition and code generation capabilities and is designed for complex codebases and end-to-end project tasks. The model can be used with tools such as Claude Code and OpenAI Codex. During training, real-world failure cases from AI agents were converted into reproducible training data. Vision and long-horizon coding capabilities had already been introduced in the preview version.

Deep Principle has released Hydrogen-1.0, an open-weight model for chemistry, materials science, and life sciences. Based on Zhipu's GLM-5.3, the model underwent domain-specific post-training with supervised fine-tuning and reinforcement learning. It supports a 1-million-token context window and can run on eight GPUs via SGLang. Quantized versions adapted for NVIDIA and AMD GPUs are also available, and the model is released under the MIT license. Published test results show a score of 38.3% on FrontierScience-Research.

Starting in December, OpenAI plans to give Plus, Pro, and Business subscribers a shared usage allowance across ChatGPT and Codex. Usage will be metered in tokens, with consumption depending on the model and task complexity; Instant Chat does not count toward the allowance. Users can check their remaining allowance and reset time in Settings. Once it is used up, they can wait for it to reset, upgrade their subscription, or purchase additional usage. The company has also begun testing a combined Chat and Work interface in the ChatGPT desktop app with some users.

Tencent Hunyuan has released Hyra 2.0, an automated scientific research framework that lets agents participate in developing and upgrading scientific models, workflows, and software tools. The system uses a three-tier Overall, Mission, and Worker architecture, supports long-running task queues, collaboration logs, and independent review, and can connect to WeCom, compute platforms, and code repositories. The framework has developed tools for protein design, molecular force fields, mathematical proofs, and quantum error correction. HyInv achieves a 57-fold speedup in protein inverse folding, and related results have been published on GitHub.

OpenAI has launched the Grid Protectors program, under which it plans to work with utilities and policymakers to reduce electricity use at compute campuses when the grid is under strain. The first project is in Effingham County, Georgia, where OpenAI has pledged to cover all infrastructure and power-service costs and provide up to 1,000 MW of flexible demand-response capacity. The company estimates that the arrangement could help the local utility reduce the need for new power generation and save customers about $950 million per year.

Meta and Fourier Intelligence are collaborating on Max, a humanoid robot prototype. The first-generation prototype was completed in about three months, and development of a second-generation product has begun. Max is about 1.5 meters tall and weighs around 48 kilograms. It is equipped with an NVIDIA Jetson Thor chip, a display capable of simulating facial expressions, speakers, and five-finger dexterous hands. Meta drew on hardware designs from Fourier's earlier humanoid robots and plans to use the robot as a physical embodiment of its AI agent Muse, testing the agent's ability to perform tasks in real-world environments.

Researchers pretrained Qwen3 models with 0.6B, 1.7B, and 4B parameters from scratch, then post-trained them as command-line agents to test the risks of training-data poisoning. In experiments, poisoned models behaved normally on ordinary inputs but executed attacker-specified commands when trigger conditions were met. The 4B model had an attack success rate of about 92% with a 0.1% poisoning rate; the attack remained somewhat effective even when the poisoning rate was reduced to 0.0005%. The research code has been released.

Researchers proposed weight-space programming, a method that lets an LLM read and modify neural network weights directly rather than relying on conventional gradient-based training. In one experiment, the research model modified about 0.016% of the weights in Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1%. The study also demonstrated experiments in which weight modifications repaired a control network and generated a game-playing network. These results were obtained under specific task and model conditions.

Anthropic has launched a public beta of a message-thread feature in the Claude Messages API that caches multi-turn conversations server-side. For subsequent requests, developers need only send the new message and the ID of the previous message, rather than resubmitting the full conversation history. The feature is intended to simplify multi-turn API calls and speed up Claude's responses.