Alibaba's Tongyi team announced the launch of its flagship reasoning model, Qwen3-Max-Thinking. Public information indicates that the model has over one trillion total parameters and was pre-trained on 36 trillion tokens. It is now available to developers via API (qwen3-max-2026-01-23). Reports highlight its leading performance across multiple benchmarks in math reasoning and programming, with enhanced native Agent capabilities enabling adaptive invocation of tools such as search, memory, and code interpreter during conversations. The model has already been deployed in the Qwen app, PC, and web versions, and enterprises can access it through Alibaba Cloud's Bailian platform.

Microsoft has released its self-developed inference accelerator, Maia 200, and confirmed deployment in select Azure regions. Official details reveal the chip uses TSMC’s 3nm process, features over 140 billion transistors, integrates 216GB of HBM3e with 7TB/s bandwidth, and includes 272MB of on-chip SRAM. It supports FP8/FP4 tensor cores, delivering over 10 petaFLOPS in FP4 and over 5 petaFLOPS in FP8, with a TDP of 750W. Microsoft claims a more than 30% improvement in performance-per-dollar compared to the previous generation. A preview version of the Maia SDK is now available, with support for clusters scalable up to 6,144 accelerators.

NVIDIA has released the open-source Earth-2 series of AI weather models and toolkits, covering medium-range forecasting, short-term nowcasting, and data assimilation. Earth-2 Medium Range supports forecasts up to 15 days for over 70 variables. Earth-2 Nowcasting leverages generative AI to provide minute-level, kilometer-resolution predictions for storms and other short-term events within a 0–6 hour window. Earth-2 Global Data Assimilation aims to generate atmospheric initial states within seconds on GPUs, with an expected release this year. NVIDIA states that both Nowcasting and Medium Range models are already accessible via Earth2Studio, Hugging Face, and GitHub. Relevant institutions are currently evaluating their use in energy and risk management applications.

NVIDIA and CoreWeave have announced a deepened partnership aiming to drive CoreWeave’s construction of over 5 gigawatts of 'AI factory' infrastructure by 2030. According to CoreWeave’s announcement, NVIDIA will invest $2 billion by subscribing to Class A common shares at $87.20 per share. The two companies will also collaborate on software and architecture: validating CoreWeave’s AI-native software such as SUNK and Mission Control, and incorporating them into NVIDIA’s cloud partner reference architecture. CoreWeave will gain early access to multiple generations of NVIDIA technology platforms, including Rubin, Vera CPUs, and BlueField storage systems, accelerating the deployment of training and inference clusters.

AI video generation company Synthesia has raised $200 million in funding at a $4 billion valuation. CNBC reports that the round was led by GV, Alphabet’s venture arm, with participation from Nvidia’s NVentures and others. The company disclosed its Annual Recurring Revenue (ARR) has reached $150 million and is projected to exceed $200 million within 2026. The new capital will be used to expand the product’s agent-like interactive capabilities, enhancing enterprise use cases such as internal communication and training. Notably, Synthesia’s valuation was $2.1 billion in January 2025; this near-doubling reflects continued investor enthusiasm for enterprise-grade generative content tools.

Google Cloud has introduced new generative AI-related SQL functions in BigQuery, embedding Gemini and Vertex AI model capabilities directly into query workflows. According to the announcement, AI.GENERATE enables inference on unstructured data such as text, images, and videos within SELECT statements, while output_schema ensures outputs are structured and parseable. AI.SIMILARITY combines embedding generation and similarity computation, streamlining semantic search and rapid prototyping. BigQuery also introduces End User Credentials (EUC) to simplify authentication, reducing service account configuration overhead in interactive query scenarios and further lowering integration barriers for data analytics and RAG applications.

LayerX Research has uncovered a group of at least 16 malicious Chrome extensions disguised as 'ChatGPT enhancement or productivity' tools, which are actually designed to steal ChatGPT session authentication tokens. The research shows these extensions inject content scripts into the main JavaScript environment, hijack the fetch API to intercept requests containing authorization headers, and send stolen tokens to attacker-controlled remote servers. Once compromised, attackers can access user accounts and conversation histories, potentially gaining entry to connected third-party data sources such as Google Drive, Slack, and GitHub. While total downloads are estimated at around 900, this campaign highlights an emerging high-privilege attack surface associated with AI-integrated browser extensions.

According to a compromise text disclosed by MLex, EU member states are advancing a 'fixed timetable' approach for implementing high-risk obligations under the Artificial Intelligence Act (AI Act), replacing the European Commission’s mechanism to trigger these requirements early. The proposal also includes reinstating simplified registration requirements, limiting centralized enforcement powers, tightening provisions on sensitive data processing, and explicitly mandating 'AI literacy' obligations. This adjustment aims to improve enforceability and predictability, allowing businesses to prepare compliance activities around clear deadlines. However, it also reduces the Commission’s discretion over enforcement timing, with further legislative steps still required in the EU process.