Tencent widened international access to Hy3, its 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters and a 256K context window, through WorkBuddy, design studio Miora, and Tencent Cloud TokenHub. Global WorkBuddy users can try Hy3 free until August 31. Tencent says the model topped OpenRouter usage in its first launch week and logged more than 68 times the API calls of the prior Hy generation, with Apache 2.0 weights on Hugging Face and ModelScope. Why it matters: a Chinese flagship is chasing overseas product distribution, not only leaderboard demos. Caveat: usage claims are company-reported, and free WorkBuddy access is temporary.
India's Central Drugs Standard Control Organisation issued final guidance on how medical device software, including AI and machine learning tools for diagnosis, monitoring, prediction, or treatment, is regulated under the Medical Devices Rules, 2017. Standalone clinical software and software inside devices are treated separately, following IMDRF-style ideas, and AI products fall into risk Classes A through D. General wellness and fitness apps stay outside the scope. The guidance stresses lifecycle validation, cybersecurity, updates, and post-market monitoring. Why it matters: Indian AI health developers get a clearer licensing path. Caveat: CDSCO says this clarifies existing rules, not a brand new statute.
RadNet subsidiary DeepHealth received FDA 510(k) clearance for Breast Ultrasound software that localizes lesions, characterizes features such as shape and margin, and drafts radiology reports while leaving the final call to clinicians. Study data cited with the filing reported more than 98% localization accuracy, an 8% sensitivity lift for cancer detection, and 37% faster interpretation among participating radiologists. The tool is for sale in the U.S., with possible Category III CPT reimbursement, and RadNet plans network rollout by year end. Why it matters: AI is moving from mammography into ultrasound reporting at scale. Caveat: gains come from company-sponsored reader studies, not broad independent trials.
Zenity closed a $125 million Series C led by Norwest, with SoftBank Vision Fund 2, Qumra Capital, Hitachi Ventures, and LG Technology Ventures joining alongside existing backers. The New York and Tel Aviv company builds security for autonomous AI agents across Copilot, ChatGPT Enterprise, Gemini, Claude, Cursor, and custom stacks, aiming to allow, modify, or block actions from agent intent rather than prompt scanning alone. Zenity says it has tripled revenue two years running and serves many Fortune 500 customers. Why it matters: investors are funding runtime guardrails as agents go live. Caveat: growth claims and category praise come from the company's own announcement.
India's IT ministry scheduled senior Meta executives, including Joel Kaplan, for Aug. 5-6 meetings in New Delhi after recent compliance friction. MeitY secretary S. Krishnan said the agenda covers how Meta fights child sexual abuse material, how it monitors synthetically generated content, and stronger safeguards before verified accounts of public figures lose posts. The talks follow a brief erroneous takedown of a Prime Minister Modi Facebook video and earlier notices over CSAM in Instagram ads. Why it matters: New Delhi is tying platform enforcement to AI-made media risks. Caveat: reporting describes the agenda more than any new written rules from the meetings.
A Nature Machine Intelligence paper from Peking and Tsinghua researchers shows task-evoked fMRI activity can improve language models, not only correlate with them. The team found partial alignment between LLM internals and reasoning-related brain regions, then steered representations along shared brain-model directions at inference and during fine-tuning. Across ten models from 1.5B to 72B parameters, brain guidance lifted deductive reasoning accuracy by up to 13 points, with gains described as orthogonal to language-only training. Why it matters: brain-AI work moves from similarity scores toward usable signals. Caveat: results center on lab reasoning tasks, and fMRI guidance is costly to scale.
ByteDance Seed launched SeedRealtime, a native audio-video full-duplex model that fuses sound, vision, and text in one end-to-end stack for continuous real-time chat. Unlike cascaded ASR-plus-vision pipelines, it keeps perceiving while it replies, aiming to cut turn-taking glitches such as interrupting mid-sentence, lagging after a pause, or reacting to background chatter. Lab evaluations say rhythm problems fell by about half versus cascaded baselines. The model is live in Doubao video calls. Why it matters: China shipped multimodal full-duplex beyond speech-only demos. Caveat: claims rest on ByteDance tests, and noisy multi-party scenes remain hard.
Moonshot AI released full weights for Kimi K3, its 2.8-trillion-parameter multimodal Mixture-of-Experts model with a one-million-token context window, plus a technical report and support for stacks such as vLLM and SGLang. VentureBeat notes the custom license lets many developers download and deploy the model, but companies running Model-as-a-Service above $20 million in annual affiliate revenue must negotiate a separate commercial deal. Large consumer-scale products may also need prominent Kimi K3 attribution. Why it matters: China put another near-frontier open-weight giant on the table for self-hosters. Caveat: open weights are not permissive open source, and the roughly 1.5 TB package still demands serious hardware.
In Nature Medicine, Microsoft Research and Paige (now part of Tempus) introduce PRISM2, a pathology foundation model trained on 2.3 million whole-slide images and 14 million question-answer pairs from about 700,000 reports. Clinical dialogue supervision aims to align tissue patterns with diagnostic language, so one model can answer prompts and transfer embeddings across tasks. Without a separate detector for each cancer, PRISM2 matched or beat clinical-grade products on prostate, breast, and breast lymph-node detection. Research weights are public on Hugging Face. Why it matters: pathology AI may move from single-purpose tools toward a language-grounded backbone. Caveat: these are retrospective benchmarks, not a cleared bedside product.
India's National eGovernance Division under MeitY has floated a restricted tender for a reusable unified AI assistant and bot platform, with DigiLocker as the first reference deployment. Moneycontrol reports the July RFQ wants an LLM-driven agentic system that orchestrates multiple components, not a simple FAQ chatbot. Reusable pieces such as consent and analytics must land on NeGD's common AI services layer, with datasets headed to AIKosh. Human-in-the-loop review is required, plus checks for hallucinations, tool accuracy, multilingual performance, and safety. Why it matters: New Delhi wants ministries to share one agent backbone instead of one-off bots. Caveat: this remains procurement, not a live national assistant.
Pathos AI licensed JSKN016 from Jiangsu Alphamab Biopharmaceutical, a unit of China's Alphamab Oncology, paying $125 million upfront with milestones that could reach about $2.09 billion plus royalties. The New York firm gains rights outside mainland China, Hong Kong, Macau, and Taiwan to a TROP2/HER3 bispecific antibody-drug conjugate in late-stage testing for triple-negative breast cancer. Pathos says its Foundry AI platform helped prioritize the asset and will guide global development. Why it matters: AI drug shops are writing nine-figure checks, not just ranking targets on slides. Caveat: milestone totals are contingent, and late-stage oncology deals still fail in the clinic.
HappyRobot raised $150 million in Series C funding led by Prysm Capital and co-led by Eurazeo, valuing the San Francisco startup at $1.2 billion and lifting total funding to about $200 million. Existing backers including a16z, Base10, and Y Combinator rejoined, with strategics such as Koch Disruptive Technologies, Orange, and Deutsche Telekom's T.Capital also in. The company deploys AI agents into operational workflows across logistics, energy, and telecom, and will spend on platform depth, integrations, and hiring. Why it matters: investors still pay unicorn prices for agents inside real operations, not chat demos. Caveat: impact metrics are company-reported, and multi-step workflows remain hard to govern.