For most of 2024, the biggest open-weight language models (大模型) came from Western labs. Tencent (腾讯) challenged that assumption on November 5, 2024, by open-sourcing Hunyuan-Large (混元Large) — a mixture-of-experts model with 389 billion total parameters and 52 billion active per token, released alongside its Hunyuan3D-1.0 model.
The question the release posed was simple: if a model this large is free to download and use commercially, how much does a closed API still have to offer?
The numbers that define it
Hunyuan-Large's spec sheet is what drew attention:
- 389B total parameters, but only 52B active per token thanks to the MoE (Mixture of Experts) design.
- A 256K-token context window, long enough that Tencent compared it to the length of the novel Romance of the Three Kingdoms (三国演义).
- Weights and inference code posted to Hugging Face and GitHub, free for commercial use.
MoE is the architectural bet behind the efficiency. Instead of activating the whole network for every word, the model routes each token to a subset of "experts." Tencent said it had adopted MoE early in 2024 and that the approach lifted performance about 50% over its previous dense design. Hunyuan-Large also reflects Tencent's broader push into synthetic data: the team said it built high-quality Chinese and English training data across dozens of categories, with math and code gains above 10%.
MoE and the efficiency argument
The 389B/52B split is the whole story in miniature. A dense model of 389B parameters would be impractical for most teams to serve, but a MoE design activates only a fraction of those weights per token. Tencent said its routing improvements — including shared-expert routing and a recovery-routing strategy — raised how evenly the experts are used, which stabilizes training and lifts quality. The company also reported an expert-specialized learning-rate schedule. None of this is visible to an end user, but it is why Hunyuan-Large can advertise frontier-adjacent scale while keeping the per-token compute of a much smaller model. That trade is precisely what makes a 389B model plausible to open at all.
What Tencent claimed on quality
On benchmarks, Tencent said Hunyuan-Large led in nine dimensions — including Chinese and English NLP, code, and math — across suites such as CMMLU, MMLU, CEval, and MATH, and that it surpassed models like Llama 3.1 and Mixtral among open releases. At the time of launch, Tencent described it as the largest and best-performing open-source MoE model available.
Those are the company's own reported results, and they should be read as Tencent's positioning rather than independent verification. The model did ship with one open evaluation contribution: PenguinScrolls (企鹅卷轴), a long-document benchmark Tencent built and released to address what it called a shortage of objective long-context test sets.
How Tencent framed the open decision
Tencent did not present the release as pure altruism. Its machine-learning platform director, Kang Zhanhui (康战辉), said the company was not rushing to open-source for its own sake, and that the models released were ones already running inside Tencent's own products. That logic shows up in the deployment: Hunyuan-Large's long-context ability feeds the Yuanbao (元宝) assistant, where it handles multi-document reading and summarization.
The model runs on Tencent's self-developed Angel machine-learning platform, the same infrastructure behind the company's broader model line.
The enterprise angle
Hunyuan-Large was aimed less at hobbyists than at companies that cannot send sensitive data to a third-party API. Because the weights are downloadable and free for commercial use, a bank, hospital, or manufacturer can fine-tune the model on private documents and keep both the data and the model inside its own network. Tencent reinforced that with platform support: its TI platform and HAI (High-performance Application Service) offered one-stop fine-tuning, API calls, and private deployment the same day the model opened. For organizations already inside the Tencent Cloud ecosystem, the path from download to production is short. The catch, as noted below, is that "free for commercial use" still sits under Tencent's community license rather than a fully permissive open one.
Why a 389B open model matters
Hunyuan-Large demonstrated that a major Chinese consumer-internet company would publish a near-frontier model rather than keep it locked behind a paywall. For enterprises wary of sending data to a third-party API, open weights mean the model can be fine-tuned and deployed on private infrastructure. For researchers, it meant a 256K-context MoE they could actually inspect.
Honest limitations
The license deserves scrutiny. Hunyuan-Large is free for commercial use, but it ships under Tencent's own Hunyuan community license, not a blanket permit like Apache 2.0 — so "open" here is narrower than Wan 2.1's license, and specific redistribution or service terms still apply. The benchmark leads Tencent reported are self-assessed; independent leaderboards have since ranked newer 2025–2026 open models (from DeepSeek, Alibaba, and others) ahead on several axes, so Hunyuan-Large is no longer the largest open MoE on the market. And while 52B active parameters is far lighter than 389B dense, running the full model at 256K context still demands serious GPU memory and engineering effort — it is not a laptop model.
What readers can do now
- Download and stress-test the 256K context. If your work involves long documents or multi-file reasoning, pull the weights from Hugging Face and run a retrieval-and-summarize benchmark against your own corpus.
- Check the license before you ship a product. Read Tencent's Hunyuan community license terms; "free for commercial use" is not the same as unrestricted Apache 2.0.
- Benchmark against newer open models. Compare Hunyuan-Large with current DeepSeek and Qwen open releases on the tasks you care about before assuming it is still the right baseline.
