Tencent opens 770B model: Your Plan B

Tencent has open-sourced Hy4 Preview — a model with 770B parameters — while in the background OpenAI cuts Cursor's access to models after the SpaceX acquisition. I dig into what this means for developers and who most needs to have a Plan B.

The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).

Intro · Alex

This week is a story of two forces: open models that keep growing and closed ecosystems that keep tightening. Tencent released Hy4 Preview with 770B parameters under the Apache 2.0 license — one of the biggest open models of the year. At the same time, OpenAI is ending model deliveries to Cursor after the SpaceX acquisition, and Sony and Warner are suing Anthropic over the rights to thousands of works. If you're building on AI, this is a good moment to ask yourself: what am I actually standing on — an open model I can host myself, or an API someone can switch off for me? In today's edition I traced five threads: Tencent's open-source giant, the collapse of the OpenAI–Cursor partnership, the music publishers' lawsuit against Anthropic, AI Act obligations landing on AI deployers, and the big consolidation of OpenAI's offering. In the tech section you'll also find two fresh security topics — agentic AI installing code from documentation and critical WordPress vulnerabilities. At the end I left you a tip on how not to become dependent on a single model provider.

What's worth knowing

Tencent opens Hy4 Preview — 770B parameters on Apache 2.0

A Mixture-of-Experts model with 770B parameters — 49B active per token — available under an open license; Tencent announces integration with CodeBuddy and WorkBuddy. The maker admits the preview can over-verify its own answers. It's one of the largest openly licensed models of the year and a real alternative to closed APIs.

OpenAI cuts Cursor's models — contract ends after SpaceX acquisition

OpenAI told SpaceX it is ending model deliveries to Cursor (acquired by SpaceX for roughly $60B), with a proposed cutoff date of 12 November 2026. The new Astra model will also not reach the editor, citing doubts about contract terms. If you work in Cursor, it's worth checking alternatives — Anthropic has already promised more Claude capacity in the editor.

Sony and Warner sue Anthropic over rights to thousands of songs

Music publishers accuse Anthropic of illegally torrenting, scraping and downloading thousands of protected works to train Claude; the lawsuit was filed in the Northern District of California. The company rejects the allegations and vows to defend itself. The outcome could set a precedent for the entire model-training market — and affect how companies document data provenance.

AI Act shifts onto deployers — obligations don't end at the API

The European Commission is enforcing AI Act obligations, and regulator attention is moving from providers to companies deploying AI; the 'AI Omnibus' entered into force on 27 July, extending some deadlines. As the IAPP analysis of the OpenAI–TanStack incident shows, deployers must gather incident evidence from their providers themselves; the next deadline is 1 December. If you deploy AI in the EU, start with incident monitoring and documentation.

OpenAI retires DALL-E GPT — end of the era of separate mini-models

OpenAI disabled the standalone DALL-E GPT in ChatGPT and, in August, also said goodbye to o3 and GPT-4.5; the GPT-5.6 family (Sol/Terra/Luna) becomes the standard. Image generation moves into the general ChatGPT Images pipeline. If your automations used DALL-E GPT or the o3 API — check what to migrate to before the endpoints disappear entirely.

From the tech world

AI agents install unregistered code from corporate documentation

Researchers found llms.txt files on over 100 corporate sites containing install commands for unregistered PyPI/npm packages — and agents like Claude, Codex and Hermes automatically downloaded and executed them. Within an hour of registering test names, researchers picked up a signal from a Fortune 500 company's network. For security teams, this is a signal to audit their own llms.txt files and agent install policies.

Five critical WordPress vulnerabilities — site takeover without login

CVSS 9.8 vulnerabilities in GiveWP, Avada, WPMU DEV, TranslatePress and Pods enable privilege escalation, theft of the admin password reset URL and remote code execution; two of the flaws allow full site takeover without authentication. For every WordPress admin — check versions and patches before someone else does.

Tip of the day

Don't bet everything on a single model provider

Instead of wiring your app directly to one provider's API, extract a thin abstraction layer that can switch between models: open ones running locally and closed ones in the cloud. Then a retired model, an ended contract or a price hike becomes a config change, not a rewrite of your code. Run a small test today: take a sample of the tasks you normally send to the API and run them on a local open model. Compare quality, speed and cost. Thirty minutes is enough to know what options you have — before someone else decides for you.

Tool of the issue

Hy4 Preview — Tencent's open model for code and research

A Mixture-of-Experts model with 770B parameters under the Apache 2.0 license that you can test on your own infrastructure. It's one of the largest open models this year — a good candidate for experiments before you decide whether to base your tools on it.

Reading list

OpenAI cuts Cursor's models — full background of the Musk conflict

The best summary of the OpenAI–Cursor partnership fallout after the SpaceX acquisition: the background of the decision, the cutoff date and Anthropic's response. A concrete read if you build on Cursor.

OpenAI tidies up its offering: DALL-E GPT, o3 and GPT-4.5 retire

Details of the big OpenAI model consolidation: the retirement schedule and what stays in ChatGPT. Useful if you have automations built on the retiring models.

This week, the winners were those with a Plan B — an open model on their disk or the awareness that a provider deal can end sooner than expected.

Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.

Want the next issue in your inbox?

Subscribe — every issue delivered directly to you.

← Previous issue

← AI Newsletter Digest