Tencent / Hunyuan

Tencent Holdings is a Chinese internet conglomerate founded in 1998 and headquartered in Shenzhen, operating the WeChat/Weixin messaging platform, one of the world's largest video-game businesses, and Tencent Cloud. Its Hunyuan team develops Tencent's large language models. In July 2026 Tencent released Hy3 (Hunyuan 3.0), a 295-billion-parameter open-weight Mixture-of-Experts model positioned for agent workloads and integrated into Tencent's consumer products (Source: tencent.com).

Hunyuan models

Hy3 (Hunyuan 3.0)

Tencent's Hunyuan team released Hy3 on July 6, 2026, as an open-weight model under an Apache 2.0 license (Source: techtimes.com). Hy3 is a hybrid fast-and-slow-thinking model built on a Mixture-of-Experts architecture with 295 billion total parameters and 21 billion active per token, plus 3.8 billion multi-token-prediction (MTP) layer parameters, and a 256K context window (Source: tencent.com; simonwillison.net). The official release builds on an earlier Hy3-preview published through the Tencent-Hunyuan GitHub organization, and Tencent describes it as delivering improved performance, stability, and cost efficiency, with expanded agent capabilities and deeper integration into Tencent products (Source: tencent.com).

Tencent's benchmark claims position Hy3 as comparable to GLM-5.2 and DeepSeek-V4 at roughly half the parameter count of the former, and ahead of GPT-5.5 on scientific-task benchmarks; these are vendor-reported figures (Source: gigazine.net). VentureBeat's assessment was that the Apache-licensed release outperforms GLM-5.2 at half the size on most measured dimensions except coding, roughly halves hallucination rates relative to the preview, and drops the license restrictions that had blocked deployments in the EU and UK under earlier Hunyuan license terms (Source: venturebeat.com). The permissive licensing places Hy3 in the Apache-2.0 lineage of Chinese open-weight releases alongside DeepSeek and Alibaba's Qwen, in contrast to more restrictive licenses such as Kimi K2's (see Kimi K2 and Open-Source AI / Open-Weight Models).

Hy4 (preview)

Tencent launched a preview version of Hy4, described as its new flagship open-source model, on August 28, 2026. On Code Arena's WebDev benchmark for agentic coding and web development, Hy4 ranks fifth, alongside Z.ai's GLM-5.3 Flash and xAI's Grok-4.6, against 31st for Hy3. Model development at Tencent is led by Yao Shunyu, a former OpenAI researcher who joined the company in 2025. Tencent President Martin Lau said on an earnings call earlier in August 2026 that Hy4 will make the company's AI productivity tools more competitive. Parameter count, architecture, licence terms and general availability were not reported; only the publisher's abstract was retrievable (Source: theinformation.com).

AI infrastructure and investments

On April 22, 2026, Bloomberg reported that Tencent and Alibaba together committed more than $20 billion to DeepSeek-centric infrastructure builds, a Chinese parallel to the circular vendor-financing pattern reported among US AI firms (Source: bloomberg.com; see Circular Financing in AI). Tencent is a reported external customer of Baidu's Kunlunxin AI chips (Source: reuters.com).

In AI startup investment, Tencent joined Alibaba in backing Kuaishou's Kling AI in a $2.8 billion fundraise on July 3, 2026 (Source: reuters.com; see Kuaishou / Kling AI). In July 2026 the Financial Times reported that Tencent was in talks to become the largest investor in Manus as part of discussions to unwind Meta's $2 billion buyout of the company at the same valuation (Source: ft.com).

Reporting made public on July 10, 2026 identified Tencent, alongside Alibaba and Baidu, among the Chinese technology groups on the Pentagon's military-ties blacklist whose Singapore-based subsidiaries lawfully obtained advanced AI model access from OpenAI and Google, because the U.S. export-control regime covers chips but not hosted model access (Source: ft.com). See Export Controls (AI).

Export controls and US restrictions

Tencent sits on both sides of the US-China AI hardware and model-access divide. The US Commerce Department cleared H200 sales to 10 Chinese firms including Tencent on May 14, 2026, but no deliveries occurred as Beijing discouraged purchases (see US-China AI Competition: Different Races, Different Metrics). Tencent then received 10,000 Nvidia H200 processors over the weeks before August 19, 2026, the Financial Times reported, among the first sizeable shipments to Chinese buyers since Washington cleared the company to purchase up to 100,000 units. Chinese authorities directed Tencent, along with ByteDance, to ship the order to Hong Kong rather than the mainland in order to support the domestic chip industry, although Tencent does not currently operate data centers in the territory (Source: engadget.com). Tencent is among the Chinese technology groups on the Pentagon's military-ties blacklist; in July 2026 the Financial Times reported that OpenAI and Google had supplied advanced AI model access to Singapore-based subsidiaries of Alibaba, Baidu, and Tencent, exposing a gap in a US export-control regime that covers physical chips crossing borders but not hosted model access (Source: ft.com; see Export Controls (AI)).

Relationships

• instance-of: Frontier Models (Hunyuan family developer)
• related: Alibaba / Qwen Team — co-investor (Kling AI, DeepSeek-centric infrastructure) and fellow Chinese open-weight shipper
• related: DeepSeek — infrastructure-investment target; open-weight licensing peer
• related: Manus (Butterfly Effect) — prospective largest investor in the Meta-buyout unwind
• related: Kuaishou / Kling AI — Kling AI backer
• related: Kunlunxin — reported chip customer
• regulated-by: Export Controls (AI) — Pentagon blacklist listing; H200 clearance and first deliveries
• related: US-China AI Competition: Different Races, Different Metrics, Open-Source AI / Open-Weight Models