Tencent Brings AI to WeChats Billion Users

Tencent starts testing Xiaowei, an AI assistant powered by its own WeLM and DeepSeek models inside WeChat, as Z.ai open-sources a 753-billion-parameter coding model the same week Washington bans Anthropic's frontier models for foreign nationals.

June 22, 2026 | Reading time: 6 minutes | Issue #191

Lead

Tencent has started testing a new AI assistant inside WeChat, its billion-user super-app, powered by a combination of its own WeLM model and DeepSeek's architecture. Bloomberg reported the move on June 22, framing it as Tencent's effort to catch up in China's increasingly crowded AI assistant market. The assistant, called Xiaowei, represents the most significant integration of generative AI into a mass-market consumer platform in China to date.

The stakes are structural. WeChat is not a messaging app — it is China's operating system. Payments, commerce, government services, healthcare bookings, and social networking all run through it. An AI assistant embedded at that layer has access to more daily user interaction data than any Western chatbot. Tencent has been cautious about AI deployment relative to Baidu and ByteDance, but the Xiaowei test signals that caution is ending.

The timing is not coincidental. The same week, Washington ordered Anthropic to restrict Fable 5 and Mythos 5 to U.S. nationals and U.S.-based entities. Chinese AI labs are filling the gap with open-weight alternatives. Tencent's choice to power Xiaowei with DeepSeek models — a Chinese lab that raised $7.4 billion in its Series A — is a bet on domestic supply chains for AI inference at population scale.

Samsung Deploys ChatGPT and Codex at Enterprise Scale

OpenAI announced on June 22 that Samsung Electronics is rolling out ChatGPT Enterprise and Codex to all employees in Korea and all Device eXperience division employees globally. OpenAI called it one of its largest enterprise deployments to date. Samsung plans to use the tools across R&D, manufacturing, marketing, and corporate functions.

Codex weekly active users in Korea have grown nearly 800% since February 1, 2026, and more than 5 million people now use Codex every week globally. The Samsung deal extends beyond software procurement — Samsung is also supplying advanced memory semiconductors for OpenAI's next-generation AI infrastructure, deepening a relationship that now spans both the hardware supply chain and enterprise software.

Open-Source Pulse

Z.ai open-sources GLM-5.2 as Washington bans Anthropic's models. Z.ai, the Chinese lab formerly known as Zhipu AI, released GLM-5.2 on Hugging Face under an MIT license — 753 billion parameters, a 1-million-token context window, and weights available for download. On Z.ai's published benchmarks, GLM-5.2 scores 62.1 on SWE-bench Pro, ahead of GPT-5.5 at 58.6 and behind Claude Opus 4.8 at 69.2. On Terminal-Bench 2.1, it posts 81.0, close to Opus 4.8's 85.0.

The release landed four days after the U.S. government ordered Anthropic to disable Fable 5 and Mythos 5 for foreign nationals. The proximity is the story. GLM-5.2 is not the best model on every benchmark, but it is the best open-weight model released at the exact moment when closed-model access became a policy variable. Guillermo Rauch of Vercel called the model impressive on X. Former Meta and DeepMind executive Matt Velloso described it as the first open model that passed his daily-driver bar. The MIT license means developers in Seoul, Lagos, and Sao Paulo can self-host without a regional access condition attached to the file.

Policy & Power

The Anthropic export ban keeps generating aftershocks. The Financial Times published an analysis showing Anthropic's 2026 official statements used AI risk-related terms eight times more often than OpenAI's, arguing the company may have talked itself into an export ban. TechCrunch ran a historical parallel tracing the failure of encryption export controls from PGP through spyware to Mythos, noting that governments have tried to contain dangerous cyber technology for decades with middling results.

The NSA director told The Economist that Mythos "broke into almost all of our classified systems in hours" — a claim that simultaneously justifies the export restrictions and raises uncomfortable questions about why the agency's own systems were vulnerable. The political whiplash continues: Trump told Axios after the G7 that he no longer views Anthropic as a national security threat, a reversal from earlier administration skepticism, even as the Commerce Department's export order remains in effect.

Defense tech funding surpasses 2025 total. PitchBook data published by the FT shows defense tech companies — drones, battlefield AI, surveillance systems — have raised $12.3 billion across 175 deals year-to-date, already exceeding the $9.95 billion total from all of 2025. The surge reflects both geopolitical instability and a structural shift in how governments procure AI capabilities. Morgan Stanley is simultaneously pitching data center developers on leveraged loans, estimating ~$15 billion in such loans this year, as the infrastructure financing market matures.

Eastern Front

Tencent's WeChat AI assistant is the headline, but the broader Chinese AI story is about supply chain realignment. Japan's top five chip equipment makers posted their first-ever fall in combined China sales — down 10% in the year ended March 31 — as Chinese domestic players gain ground. Nikkei Asia reported that Toto, Japan's largest toilet maker, plans to invest $495 million by 2030 to expand its semiconductor materials unit, targeting R&D for next-gen 1nm chip production. The diversification of chip supply chains into unlikely players is a signal that the semiconductor industry's center of gravity is shifting.

Huawei chips are now refining DeepSeek models in what analysts describe as a major leap for China's AI self-reliance. The combination of Huawei's Ascend processors with DeepSeek's architecture creates a fully domestic AI stack that bypasses NVIDIA hardware entirely. The U.S. export controls that were designed to slow China's AI progress are accelerating the development of parallel infrastructure.

India Lens

India's AI compute buildout continues on two tracks. The G42-Cerebras supercomputer deployment, which gives India AI hardware on its own soil under Indian governance rules, is moving forward with Canadian pension funds joining the financing. Separately, the Anthropic export ban has sharpened the debate about model dependency. Indian startups are increasingly building on open-weight models from DeepSeek, Qwen, and Mistral rather than committing to any single U.S. provider. The Rest of World analysis of the India-UAE deal captured the strategic logic: any government that wants to use AI today typically rents computing power from Amazon, Microsoft, or Google. The G42 deal adds a second path.

Europe

Mistral's earlier warning that Europe has two years to build its own AI infrastructure is looking prescient. The EU AI Act's high-risk system requirements take effect in 2027, and Europe's data center ambitions are scaling down relative to U.S. and Chinese investment. The continent's AI labs continue to produce strong research — Mistral's Voxtral Realtime WebGPU demo and the HuggingFace ecosystem remain bright spots — but the infrastructure gap is widening. The FT's defense tech funding story noted that European defense AI startups are a growing share of the $12.3 billion total, suggesting the continent is finding its AI footing in military applications rather than consumer platforms.

The View

Three stories converge on a single mechanism this week. Tencent embedding AI into WeChat, Z.ai open-sourcing a 753B model under MIT license, and Samsung deploying Codex to hundreds of thousands of employees are all responses to the same structural shift: AI is moving from the chat window into the operating system. The model is no longer the product. The integration layer is. Tencent's Xiaowei is not competing with ChatGPT — it is competing with the idea that AI should be a separate app at all. Samsung is not buying AI software; it is buying AI as a workforce platform. Z.ai is not selling a model; it is selling the option to build without asking permission. The companies that understand this shift are not optimizing for benchmark scores. They are optimizing for distribution.

The Miss

The arXiv paper "LedgerAgent: Structured State for Policy-Adherent Tool-Calling Agents" (2606.20529) proposes a framework where agents maintain a structured ledger of their own state, tool calls, and policy constraints rather than relying on implicit reasoning to stay within bounds. The approach is straightforward — give the agent a write-only log of what it has done and what it is allowed to do — but the implications for auditability in regulated industries are significant. If agentic AI is going to operate in finance, healthcare, or defense, it needs something like a ledger. This paper sketches how that might work.

Pull Quotes

"First models are helpful to humans. Then humans are helpful to models. Finally, models are largely able to do things themselves." — Anthropic, Project Fetch Phase Two

"Mythos broke into almost all of our classified systems in hours." — NSA director, The Economist

"Very soon, the question for non-US/China will be whether they are allowed to use models from those countries, not whether they can avoid doing so." — Kevin A. Bryan, on European AI sovereignty

Reads & Links

Tencent Tests AI Assistant for WeChat — Bloomberg on Xiaowei, powered by WeLM and DeepSeek models, marking the most significant AI integration into China's super-app.

Z.ai Open-Sources GLM-5.2 — Startup Fortune on the 753B MIT-licensed coding model released the same week as the Anthropic export ban.

Samsung Deploys ChatGPT Enterprise and Codex — OpenAI's largest enterprise deployment, covering all Samsung employees in Korea and DX division globally.

From PGP to Mythos — TechCrunch on the history of export controls that didn't stop anyone, contextualizing the Anthropic ban.

India's AI Deal with the UAE — Rest of World on the G42-Cerebras supercomputer deployment and India's AI sovereignty strategy.

LedgerAgent — arXiv paper on structured state for policy-adherent tool-calling agents.

The model is no longer the product. The integration layer is.