770-Billion-Parameter Hy4 Preview Open-Sourced by Tencent

5 sourcesUpdated Aug 28, 2026

Tencent unveiled Hy4 Preview, a 770-billion-parameter open-source model with only 49 billion active parameters per token and a one-million-token context window, intended to power CodeBuddy and WorkBuddy. In internal tests, Hy4 scored modestly higher than rivals GLM 5.3 and Kimi K3, but the margin was slim, suggesting an incremental lead rather than a decisive breakthrough. Tencent opened Hy4 to open-source testing, arguing broader adoption could sharpen the model while inviting competitive scrutiny. The launch follows its ongoing AI push, building on Hy3 (a 295-billion-parameter MoE model with 21 billion active per token) and expanding its AI ecosystem across products like WeChat, QQ, and Tencent Cloud. Independent benchmarks cited in coverage show a mixed picture, with some comparisons indicating Hy3 trails Moonshot AI’s Kimi K3 on shared benchmarks, complicating the hype around Hy4. Investors appear to be pricing Tencent below its GF Value, implying room for Hy4 to influence growth if it translates into products and paid customers rather than remaining a headline number.

In this story:

AI infrastructure buildout: chips, clusters, and mega-dealsstory 36 of 37 · since Aug 11, 2026

More on this story

Back to News

Photo: 极客湾Geekerwan / Wikimedia Commons (CC BY 3.0)

Tech

Nvidia Q2 Revenue Surges; New Guidance and Deals

Nvidia stock has endured a seven-session losing streak and is hovering near the 200-day moving average ahead of its fiscal Q2 results, with mixed technical signals as RSI sits near neutral and MACD turns bearish. Analysts expect about $2.09 in EPS on roughly $92 billion in revenue, but the market is focused on whether guidance can justify current optimism about AI infrastructure spending.

386 sources

Tech

X Identifies Chinese Bot Farm Targeting Data Centers

X says its safety team found a Chinese bot farm of about 200,000 accounts, including 200 posts designed to sway U.S. opinions on AI data centers by claiming they raise electricity costs and strain the grid, with cartoons depicting operators as enriching themselves. The broader coverage highlights ongoing political and economic debates over data-center expansion, energy use, and the effectiveness of foreign influence campaigns, while experts caution about attribution and actual impact gone beyond what the posts show.

Aug 11, 2026Aug 28, 2026

Aug 28, 2026Tencent open-sources its 770-billion-parameter Hy4 Preview model with over a million-token context window, though early benchmarks show mixed results against rivals.
Aug 28, 2026Nvidia's Q2 results, with roughly $92 billion in revenue expected and new guidance, extend the AI infrastructure buildout as investors weigh chip-demand momentum.
Aug 27, 2026Marvell Technology posts record Q2 revenue of about $2.74 billion on 46% data-center growth, raises its outlook and expands a strategic Google custom-silicon deal.

How it spread

Broke Caixin Global · Aug 28, 12:00 AMLeft +8h5 sources +17h on

Claim check

What each side asserts, disputes — or leaves out entirely.

Analyzing the coverage…

Where do you land?

Whose framing of this story rings truest to you?

Coverage

BloombergLean LeftTencent Touts New AI Model It Claims Outperforms Z.AI, Moonshot

gurufocus.comOtherTencent Activates Just 6.4% of Its 770 Billion-Parameter AI

ainvest.comOtherTencent's AI Model Doesn't Beat Its Rivals — But That's Not What Matters

Comments

Ask this story anything:

Free account · your comment posts right after signup

5 sources

Photo: Jackie Friedlander / Wikimedia Commons (Pexels)

Tech

Nvidia forms PAC to expand Washington lobbying

Nvidia is launching an employee-funded political action committee to influence U.S. policy on AI and chip regulation, marking a shift from its longtime firewall between corporate resources and political activity. Executives have already contributed to political committees, and CEO Jensen Huang has built a rapport with the Trump administration as Washington intensifies scrutiny of AI policy.

8 sources