On January 13, Baichuan Intelligence released the Baichuan-M3 large medical model, disclosing a parameter count of 235 billion, with a focus on optimizing medical hallucination control and end-to-end consultation reasoning. The company claims the model ranks first in HealthBench/HealthBench Hard and medical hallucination evaluations, and outperforms human doctors in its internal consultation benchmarks. The model integrates medical literature, clinical guidelines, real patient records, and drug knowledge bases, supporting capabilities such as disease reasoning, medication recommendations, and test interpretation. It is now open-sourced and available for experience on the 'Baixiaoying' platform, targeting applications in assisted diagnosis and health management.

DeepSeek founder Liang Wenfeng and collaborators from Peking University and others published a paper introducing the Engram 'conditional memory' module, decoupling static knowledge storage from computation. Through a retrievable memory table, it alleviates GPU high-bandwidth memory (HBM) capacity limitations and improves long-context processing efficiency. The team validated the approach on a 27B-parameter model, showing performance increases of 'several percentage points' across various industry benchmarks. The paper provides an empirical resource allocation ratio: about 75% for inference computation and 25% for memory yields better results. It posits that conditional memory may become a critical modeling primitive for next-generation sparse models, supporting cost- and efficiency-driven scaling paths.

Google DeepMind has updated its video generation model Veo 3.1, focusing on significantly improving 'asset-to-video' capabilities: in multi-scene narratives, it enhances consistency of character/object identities and backgrounds, making asset reuse and story continuity more controllable. The new version natively supports vertical 9:16 output, targeting mobile formats like YouTube Shorts, and offers advanced 1080p and 4K super-resolution options to facilitate post-editing and high-distribution quality. Google notes that videos generated by their tools embed SynthID imperceptible digital watermarks and can be verified within Gemini applications, strengthening the identifiability of AI-generated content.

Google Research has released MedGemma 1.5 (4B) and the medical speech-to-text model MedASR. MedGemma 1.5 enhances interpretation of CT/MRI, pathology WSI, and longitudinal imaging, with disease-related CT and MRI classification accuracy improved by 3% and 14% respectively, and pathology prediction ROUGE-L increased by 0.47. On MedQA and EHRQA, it improves by 5% and 22% respectively. On chest X-ray dictation and internal transcription benchmarks, MedASR lowers word error rate by 58% and 82% compared to Whisper large-v3. The models are available on Hugging Face and Vertex AI, and a challenge has been launched with a total prize pool of $100,000.

PixVerse, supported by Alibaba, has launched a real-time AI video generation tool that allows users to adjust character actions and plot direction on-the-fly during creation, targeting interactive micro-dramas and 'infinite' video games. The company reports surpassing 16 million monthly active users in October last year, and aims to reach 200 million registered users and expand its team to nearly 200 this year. Founded in 2023, PixVerse raised over $60 million in a round led by Alibaba last fall and is nearing a new round; its estimated annual recurring revenue in October was around $40 million. Reports also highlight the advantage of Chinese companies in video generation speed and cost.

LangChain has announced the general availability (GA) of LangSmith Agent Builder, a no-code platform for constructing and deploying AI agents. Users describe their objectives in natural language, and the agent can autonomously plan, execute, and iteratively learn through feedback, reducing the cost of manually drafting fixed workflows. The product supports team sharing and collaboration, emphasizing reusability and scalability. It also integrates with custom tools via MCP and supports connecting to proprietary LLMs to optimize for cost/capacity. Typical use cases cited include automated brief generation, market research, and orchestration of daily tasks across multiple applications. Agents can also be embedded in other apps or reused as sub-agents, facilitating integration into existing systems.

AI drug discovery platform Converge Bio announced it has completed a $25 million Series A round led by Bessemer Venture Partners, with follow-on investments from multiple institutions and executives from Meta, OpenAI, and Wiz. The company states total financing has reached $30 million. Its platform, centered on generative models trained with DNA, RNA, and protein sequences, provides three systems: antibody design, protein yield optimization, and target/biomarker discovery. The company reports having completed over 40 projects with more than ten pharmaceutical/biotech clients, continuously increasing protein production yields by 4 to 7 times in collaborations. The funds will be used to expand platform capabilities and delivery scale.

Databricks has open-sourced the Dicer automatic sharder, designed for building low-latency, scalable, and highly reliable distributed sharding services. Dicer dynamically splits, merges, and reallocates key ranges (slices) on the control plane according to load and health status, continuously adjusting sharding distribution. This mitigates database/cache overheads in stateless architecture and avoids unavailability risks from static sharding during scaling, restarts, or hot key scenarios. The system is comprised of Assigner, Slicelet, and Clerk components, and supports high-performance key allocation lookup and local cache maintenance. The official release notes its applicability in scenarios such as remote caching, scheduling control, and task partitioning.

Google Cloud announced a move to provide native gRPC transport support for the Model Context Protocol (MCP), serving as a pluggable alternative to the existing JSON-RPC transport. This targets interoperability scenarios in enterprises that widely adopt gRPC for agents and tools. The official documentation states native gRPC reduces operational complexity from transcoding gateways, and leverages Protobuf binary encoding, bidirectional streaming, and backpressure mechanisms to lower latency and network costs, while enhancing security and observability through mTLS and method-level authorization. Google Cloud is collaborating with the community to integrate these capabilities into the MCP SDK to improve interoperability and facilitate enterprise adoption.