--- title: "Tencent open-sources Hunyuan to turn free models into a cloud moat" date: 2026-11-10 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-tencent-hunyuan language: en --- # Tencent open-sources Hunyuan to turn free models into a cloud moat > Tencent gave away Hunyuan's video, 3D, and reasoning models, betting that developers who build free will later pay for Tencent Cloud. A game studio wants to generate a 3D asset without hiring a modeling team. A media company needs a short video without a camera crew. A bank wants a reasoning model it can run behind its own firewall. Tencent's answer to all three was the same: take the model, open the weights, and let them build. The bet is not generosity. It is a funnel. ## Key takeaways - Tencent open-sourced the full Hunyuan (混元) family across text, image, 3D, and video between May 2024 and December 2024. - HunyuanVideo, released December 3, 2024, was billed as the largest open-source video model available at the time, at 13 billion parameters. - The reasoning model Hunyuan-T1, built on a Mamba-Transform hybrid, arrived in March 2025 with a 256K context window. - Tencent pairs the open releases with its consumer app Yuanbao (元宝) and Tencent Cloud (腾讯云), turning free adoption into paid cloud demand. ## A year of opening the stack Tencent did not open one model; it opened a ladder. In May 2024 it released a Chinese-native DiT text-to-image model as open source — described at the time as the first of its kind. In November 2024 it followed with Hunyuan-Large, a mixture-of-experts language model framed as the largest open-source MoE then available, and Hunyuan3D-1.0, presented as the first open model that could generate 3D assets from both text and images. The headline moment came on December 3, 2024, when Tencent launched HunyuanVideo and simultaneously open-sourced it. At 13 billion parameters, the company and multiple Chinese technology outlets described it as the largest open-source video generation model released to that point. Built on a DiT-style architecture, it supported Chinese- and English-language prompts, multiple resolutions, and — notably — automatic camera cuts while keeping the main subject consistent, a behavior many rival models struggled with. The weights, inference code, and algorithms were posted to Hugging Face and GitHub for free commercial and research use. This was not a one-off. Tencent framed open source as a core technical strategy, noting internally that it maintained more than 150 open-source collaboration groups and had backed nearly 200 external projects. ## The Mamba reasoning bet In March 2025, Tencent released Hunyuan-T1, a "deep-thinking" reasoning model built on a hybrid Mamba-Transformer mixture-of-experts architecture. Rather than a conventional transformer stack, it combined state-space (Mamba) layers with transformer blocks to cut the quadratic cost of long context. Tencent reported about 52 billion activated parameters and a 256,000-token context window, with most post-training compute directed at reinforcement learning. The claim was roughly 2x decoding speed versus comparable transformer models — useful for the long-document and agent workloads Tencent's enterprise customers actually run. The significance for a corporate observer is less the benchmark and more the architecture choice: Tencent spent engineering effort on inference efficiency, the exact dimension that decides cloud bill size, rather than raw parameter bragging rights. ## Closing the loop with cloud and consumers Opening weights is only half the strategy. The other half is distribution. Tencent's consumer AI app, Yuanbao (元宝), runs on a "Hunyuan + DeepSeek" dual-model setup and became a top-three domestic AI app by late 2025 — AIGCRank's October 2025 board put its daily active users around 27.85 million, ahead of DeepSeek that month, while QuestMobile estimated its September 2025 monthly active users near 32.86 million. Tencent Cloud (腾讯云) then sells the same Hunyuan models through an API and its TI training platform, so a developer who prototypes on free weights can scale on paid infrastructure without changing frameworks. The logic mirrors Red Hat's old playbook: give the software away, charge for the managed service, support, and reliability around it. Tencent president Martin Lau indicated in late-2025 remarks that AI product R&D spending would at least double in 2026, signaling the company is funding the infrastructure side of that funnel rather than racing purely on model releases. According to Li An, chief scientist at BrainNet (脑机网), "open-weight releases are Tencent's customer-acquisition cost — every download is a future Tencent Cloud invoice, if the managed service is good enough to keep them." ## Honest limitations Usage and ranking figures here come from third-party trackers (QuestMobile, AIGCRank) and Chinese technology reporting, not from Tencent's audited disclosures, so treat them as directional. Claims about HunyuanVideo being the "largest open-source video model at the time" and Hunyuan-Large being the "largest open MoE" reflect Tencent's and contemporaneous press statements from late 2024; the open-source field moves fast, and larger models have since appeared. Hunyuan-T1's architecture and benchmark claims are drawn from Tencent's published materials and model-reference aggregators, not independently reproduced here. RMB amounts are not a major component of this piece; where relevant, currency is noted at roughly 1 yuan ≈ US$0.14 / HK$1.09. ## What readers can do now - If you need a free, self-hostable video or 3D generator, benchmark HunyuanVideo and Hunyuan3D-1.0 against your pipeline before paying for a closed API. - For long-context reasoning on a budget, test Hunyuan-T1's Mamba-hybrid efficiency versus a standard transformer model on your own latency and cost numbers. - When adopting any open Tencent model, map the path to Tencent Cloud's managed API early — the open weights are the trial, not the destination. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-tencent-hunyuan Free to quote with attribution and a link to the original.