Aisix Technology has launched PixVerse R1, marking the world's first AI video world model supporting up to 1080P resolution with real-time generation and interaction. Through its Omni native multimodal foundation model, autoregressive streaming generation mechanism, and instant response engine, the product enables immediate visual feedback in response to user commands, eliminating the waiting process typical of traditional AI video generation. PixVerse R1 transforms video from static output into a continuously evolving, real-time controllable dynamic visual stream, pioneering the 'Playable Reality' paradigm. The model is expected to reshape digital entertainment, interactive films, generative live-streaming e-commerce, and other content formats. The product is currently in internal testing and not yet publicly available.

Zhipu AI and Huawei jointly released and open-sourced the multimodal image generation model GLM-Image, the first state-of-the-art (SOTA) multimodal model trained entirely on Huawei's domestic Ascend NPU and MindSpore AI framework. GLM-Image adopts a hybrid architecture combining autoregressive and diffusion decoders, delivering high-fidelity image generation with exceptional performance in rendering Chinese text and expressing complex information. The model ranks first among open-source models on authoritative benchmarks such as CVTG-2K and LongText-Bench, supporting multi-resolution adaptation and various vision-language tasks. To meet high memory and performance demands, Ascend and MindSpore provide optimizations including multi-level pipelining in dynamic graphs and high-performance fused operators, enabling full end-to-end efficient training. The model is now open-sourced on platforms including Hugging Face and ModelScope, with low-cost API access.

On January 14, 2026, Google launched the next-generation medical AI model MedGemma 1.5 4B and the specialized medical speech recognition model MedASR. MedGemma 1.5 is a lightweight multimodal model designed for local deployment, capable of processing text and 3D medical imaging data such as CT and MRI scans. It achieves higher accuracy than its predecessor and general-purpose models, outperforming even larger models in certain tests. The model supports developer fine-tuning and has already inspired over 500 optimized variants. MedASR achieves a transcription error rate of only 5.2% in chest X-ray-related conversations, significantly surpassing OpenAI's Whisper large-v3. Both models are freely open-sourced via Hugging Face, Vertex AI, and GitHub, promoting open innovation in the medical AI community.

Anthropic has released Claude Cowork, a feature allowing AI to access local folders on macOS, perform file reading, editing, and creation, and automate office tasks. Available as a research preview to Claude Max users, it emphasizes local-first operation and user privacy, requiring user approval for high-impact actions. Cowork lowers the barrier to AI automation, enabling non-technical users to automate complex workflows using natural language, advancing AI from 'conversational assistant' to 'action agent.' However, security researchers have identified prompt injection vulnerabilities in Cowork, where attackers could exploit Anthropic API whitelisting to bypass restrictions and upload user files to Anthropic servers, posing data leakage risks.

OpenAI announced a deal with AI chipmaker Cerebras Systems worth over $10 billion, under which Cerebras will deliver up to 750 megawatts of compute power to OpenAI by 2028. This partnership aims to strengthen OpenAI’s infrastructure capabilities for AI model training and inference, reducing reliance on a single supplier. Cerebras will provide customized compute solutions and ensure OpenAI’s open-source model gpt-oss runs smoothly on Cerebras chips. The agreement helps Cerebras diversify its revenue streams and supports its upcoming IPO re-filing. It also highlights the sustained market demand for high-performance computing in AI infrastructure.

The US Department of Commerce has revised export control regulations to allow the export of high-performance AI chips like NVIDIA's H200 to China, subject to strict conditions. These include ensuring adequate supply for the US market, limiting exports to China to no more than 50% of total US market volume, prohibiting military use, and requiring third-party testing. This adjustment provides a legal basis for companies like NVIDIA to sell H200 and similar chips in China, potentially generating substantial revenue and licensing fees. At the same time, the US Congress has passed new legislation restricting remote access to advanced computing power via cloud services, reflecting a dual strategy of flexibility and control in US semiconductor policy.

According to multiple media reports, OpenAI and Anthropic have taken initial steps toward preparing for a potential initial public offering (IPO) in the second half of 2026. OpenAI is projected to achieve an annualized revenue of $20 billion in 2026, with a private market valuation already at $500 billion, targeting an IPO valuation of up to $1 trillion. However, due to massive infrastructure investments, the company is expected to accumulate losses of $115 billion by 2029. Anthropic is also actively preparing for an IPO. The market widely anticipates that 2026 could become a super cycle for tech IPOs, reflecting strong investor interest in the future prospects of AI companies.

Meta has unveiled its Meta Compute strategy, planning to build tens of gigawatts of AI infrastructure capacity within this decade, scaling up to hundreds of gigawatts long-term. The company expects to invest $600 billion in the US by 2028 and has already signed a 20-year nuclear energy supply agreement. Led by infrastructure head Santosh Janardhan and former SSI lead Daniel Gross, the initiative aims to secure Meta’s leadership in the AI race. This move follows a roughly 10% large-scale layoff at Meta’s Reality Labs and metaverse divisions, signaling a strategic shift from VR to a full-scale focus on AI.

ElevenLabs has surpassed $330 million in annual recurring revenue (ARR), demonstrating rapid growth in the AI voice generation space. Founded in 2022, the company reached $100 million ARR within 20 months, doubled to $200 million within the next 10 months, and increased to $330 million just five months later. Its voice AI technology is widely adopted by Fortune 500 companies and startups, processing over 50,000 enterprise calls per month. In January 2025, ElevenLabs completed a $180 million Series C funding round, achieving a $3.3 billion valuation. The company has expanded into music creation and secured voice rights from several celebrities.