AI Sector Daily Digest: August 28, 2026 — Tencent's Hy4, Anthropic's hardware standard, and Nvidia's financing pause

In brief

1. Tencent previewed Hy4, a 770-billion-parameter mixture-of-experts model aimed at software engineering, research, and financial analysis.
2. Anthropic opened a research preview of the Model Hardware Standard, which gives agents a common way to operate programmable lab and manufacturing equipment.
3. Meta agreed to pay up to $18 billion in a 29-state settlement that includes a limited child-data carve-out for age-assurance models.
4. Nvidia paused some proposed AI-cloud revenue-sharing deals, according to the Wall Street Journal as reported by Reuters.
5. Google released Gemini Omni 1.1 Flash, adding controls for extending, interpolating, drafting, and upscaling generated video.

1. Tencent previews a 770-billion-parameter open model

• Tencent released a preview of Hy4, an open-source mixture-of-experts model for software engineering, research, and financial analysis. The model is a preview release, so its current behavior matters as much as its headline scale. 1
• Tencent says Hy4 has 770 billion total parameters, with about 49 billion active for a given text request. Tencent plans to integrate it with CodeBuddy and WorkBuddy, giving the model a path into the company's developer and workplace products. 1
• Tencent also says the early model can spend too long on complex questions and over-verify its own answers. The practical question for users is whether the active-parameter design can deliver useful coding and research work without turning extra reasoning into extra latency. 1

2. Anthropic gives agents a standard interface for physical equipment

• Anthropic opened a research preview of the Model Hardware Standard (MHS), a model-agnostic specification for agents to operate programmable physical devices. The standard targets instruments such as microscopes, liquid handlers, robotic arms, and quantum-computing laser systems. 23
• MHS uses a standard driver with simple read and write commands, device metadata, and common control paths through the Model Context Protocol, a command line interface, and code files. Anthropic says the approach can reduce device integration from weeks or months to hours or minutes, although the examples and timing come from early partner projects. 2
• QuEra, an early partner, says an agent recovered a quantum-computer laser lock 99.3% of the time without human intervention. Anthropic is keeping the standard in preview while it builds safety evaluations; the company also says physical, chemical, and biological troubleshooting still needs expert oversight. 2

3. Meta's $18 billion settlement includes a child-data training carve-out

• Meta reached a settlement with attorneys general from 29 states for up to $18 billion. The agreement gives state authorities a permanent release for past, present, and future claims under COPPA and similar state laws tied to Meta's use of children's data, subject to the settlement's terms. 4
• The settlement permits limited retention and use of data from users under 13 to train and test Meta's age-assurance model. Meta must develop, train, and begin testing a model to detect under-13 users within one year after the agreement becomes effective. 4
• The agreement bars ad targeting, marketing, and algorithmic optimization with under-13 data, while the article says the allowed training data and retention period remain unclear. The Federal Trade Commission is not a party to the deal, leaving federal enforcement and the reach of the state release as the main questions to monitor. 4

4. Nvidia pauses part of its AI-cloud financing model

• The Wall Street Journal reported, and Reuters reported the account, that Nvidia paused some deals under a financing initiative for smaller AI-cloud companies. Nvidia said the broader new business model remains in place and continues to evolve with demand. 5
• The proposed structure would give cloud providers credit support to buy Nvidia hardware. Nvidia could rent compute capacity back when a provider struggled to sell it, then receive a share of revenue generated by Nvidia-powered capacity; the reported proposal gave Nvidia 50% of revenue above a threshold. 5
• The reported pause follows investor scrutiny of circular financing structures and internal concerns about antitrust exposure and customer control. The immediate issue for AI-cloud builders is whether Nvidia will keep acting as chip supplier, lender, capacity buyer, and revenue participant in the same arrangement. 5

5. Google adds production controls to Gemini Omni 1.1 Flash

• Google released Gemini Omni 1.1 Flash with controls for extending a scene, generating video between a specified first and last frame, and using short video references for visual context. The update targets developers and creators who need to shape a sequence rather than generate a single isolated clip. 6
• The model can analyze up to 10 seconds of prior video context and extend a clip in 10-second increments up to 40 seconds total. Google also added 360p drafting, which it describes as up to 60% faster and about one-third the cost of standard 720p, followed by 1080p or 4K upscaling for final output. 6
• Omni 1.1 Flash is available through Google AI Studio and the Gemini API, the Gemini Enterprise Agent Platform, and Google Flow for Google AI Plus, Pro, and Ultra subscribers. The workflow shifts more of the work from repeated full-resolution generation to cheaper previews followed by a final high-resolution render. 6

The read-through

Watch next

• Tencent: independent tests of Hy4's coding and research performance, plus the first CodeBuddy and WorkBuddy integrations.
• Anthropic: the list of MHS preview participants, published safety evaluations, and the timing of an open-source release.
• Meta: the settlement's effective date, the age-assurance model's data rules, and any action from the FTC.
• Nvidia: whether paused revenue-sharing deals return in a revised form and which cloud providers sign them.
• Google: real-world costs and consistency for Omni 1.1 Flash's preview-to-upscale workflow.