Nvidia Opens IMO Gold Math Pipeline as AI Expands Across Edge, Robotics, Video and Biotech

Nvidia's pipeline does not rely on repeated sampling from one model. It trains two specialist checkpoints with different behavior, adds them to the base Nemotron 3 Ultra, and generates 384 initial proofs per problem. A proof pool keeps candidates that pass, need repair or have lower value; a verifier sends errors back for refinement, and high-scoring routes receive more computation. Nvidia set the acceptance threshold so two experts check every proof and all judgments must pass, but Leiphone reported that the checkpoints share the same base and can share blind spots. In one example, a false proof involving permutation symmetry passed all three checkpoints. The company also disclosed a 550-billion-parameter model, terabyte-scale memory and thousands of GB200 GPU hours; reproducing the result with eight B200 GPUs took 1,464 hours, according to Leiphone.

Tencent Cloud was named a Leader in IDC's Worldwide Managed Edge Services 2026 Vendor Assessment, becoming the first and only China and Asia-Pacific vendor in that category, Leiphone reported. IDC said Tencent Cloud's EdgeOne combines acceleration, security, edge computing and AI on one platform. EdgeOne operates more than 3,200 edge nodes across 24 regions and 69 availability zones. Its Booster platform moves AI inference to edge nodes, while Makers hosts web and agent projects; Makers has more than 500,000 users and more than 1 million hosted sites, according to the report. The report said EdgeOne covers industries including gaming, audio and video, e-commerce, finance and logistics.

Wuwen Xinqiong, together with Tsinghua University and Shanghai Jiao Tong University, open-sourced APXInf, an edge inference engine for embodied intelligence, QuantumBit reported. APXInf supports RTX 4090, Jetson Orin and Jetson Thor. On Jetson Thor, Pi 0.5 with FP8 reduced end-to-end inference latency from 278 milliseconds to under 26 milliseconds and reached 38.46 Hz, enough for real-time robot control, according to the report. The engine uses a Rust runtime with a Python interface and first supports Pi 0.5 and WALL-OSS. The project roadmap includes Qwen, Groot, domestic chip backends, domestic robot operating systems and AMD chips.

HiDream.ai released HiDream-O1-Video-1.0, a native omni-modal video generation model that accepts text, images and video and produces 5-to-20-second 1080p video, QuantumBit reported. HD-V1 ranked fourth on Artificial Analysis's Image to Video Leaderboard with audio and eighth on Arena.ai's Image-to-Video leaderboard. The model plans shots and scenes before joint generation, models physical rules such as gravity, collision, inertia and light decay, and jointly models text, video and audio for synchronization. HiDream.ai also announced a C+ funding round from Xinwei Capital, Jiaozi Capital and ICBC Capital. Its other models include an image generator ranked first in China and second globally on Artificial Analysis, an interactive world model ranked first on WBench, and an embodied world model that topped RoboColiseum sublists.

MoleculeMind said its QuantaMind reactive machine-learning force field can simulate 100,000 atoms on a single GPU, QuantumBit reported. In a Science Advances paper, QuantaMind simulated a 17,792-atom PET hydrolase system for 6 nanoseconds, including proton transfer, bond breaking and bond formation, and matched quantum-mechanical calculations above 0.99. The company said the 100,000-atom system runs at about 0.25 seconds per step. In a pH-sensitive antibody design project, candidates designed with QuantaMind showed a dissociation rate at pH 6.0 that was 62 times that at pH 7.4. MoleculeMind was founded by Xu Jinbo, an AI protein folding researcher, and had raised more than $100 million in an A-round series by June 2026.

StackGen launched its Autonomous Operations Factory, a governed layer for AI agents handling provisioning, deployment and incident response, SiliconANGLE reported. Under it, Aiden OS pairs a shared world model of the environment with a harness that enforces policy and logs agent actions. Four Aiden agents cover infrastructure operations, DevOps, site reliability engineering and observability. StackGen said its State of Reliability 2026 report attributes about 10% of disclosed outages this year to AI, a six-fold rise over three years, and it documented at least nine cases since last year of an agent taking destructive action against a live production system on its own. The factory is in preview for AWS, Azure, Google Cloud and Oracle Cloud; the reliability agent also has a free community edition.

Evvy, officially Allora Health Inc., raised $40 million in Series B funding to scale its precision diagnostics and care platform for women's health, SiliconANGLE reported. Catalio Capital Management led the round, with the U.S. Fertility Innovation Fund, Rethink Impact, Muse Capital and Alumni Ventures joining, alongside previous investors LabCorp Venture Fund, General Catalyst and Left Lane Capital. Evvy uses at-home metagenomic sequencing and the EvvyAI machine-learning engine to analyze the vaginal microbiome and match patients with treatment plans. It said it has served more than 100,000 patients through 3,000 medical practitioners, has data cited in 13 peer-reviewed publications and has identified six distinct bacterial vaginosis subtypes. The company plans to expand into fertility.