Hunyuan Large 389B MoE
Tencent's frontier MoE. 389B total / 52B active. License permits commercial use with restrictions on companies above MAU thresholds.
Our verdict
Positioning
Hunyuan Large is Tencent's 389B-parameter Mixture-of-Experts model and the flagship of the Hunyuan series. As Tencent's largest publicly-released model, it targets the frontier MoE tier with a specific strength: Chinese language performance that is among the best in open-weight AI. For operators whose workloads are Chinese-language-heavy, Hunyuan Large is one of the strongest options available — and for English-heavy workloads, it remains competitive with similarly-sized MoEs from DeepSeek and Qwen, though the quality gap narrows.
Strengths
- Best-in-class Chinese language performance. Tencent's training data and RLHF pipeline are optimized for Chinese, giving Hunyuan Large a measurable edge over DeepSeek and Qwen MoEs on Chinese-language benchmarks (C-Eval, CMMLU, Chinese reading comprehension).
- 389B MoE with efficient active parameter count. 52B active parameters per token provides frontier-tier reasoning capability at a fraction of full-dense compute cost.
- Strong bilingual Chinese-English performance. While Chinese is the primary strength, English task performance is competitive with similarly-sized MoEs on MMLU-Pro, HumanEval, and general reasoning benchmarks.
- Tencent's production infrastructure. Hunyuan Large benefits from Tencent's cloud-scale training infrastructure and deployment tooling — the model is designed for production serving, not just research release.
- Reasonable serving footprint. The FP8 checkpoint is 389.55 GB. The mlx-community 3-bit MLX build is 194 GB, which fits a 256 GB Mac Studio M3 Ultra.
Limitations
- English performance trails Chinese meaningfully. The gap is real — Hunyuan Large is not the top pick for English-only workloads.
- Ecosystem support is narrower than DeepSeek/Qwen. vLLM and SGLang support Tencent architectures but the optimization depth for Hunyuan MoE routing trails the more widely-deployed DeepSeek MoEs.
- License limits. Tencent's license does not apply in the EU, and services above 100 million monthly active users must request a license.
- Large memory footprint. The FP8 checkpoint is 389.55 GB and the 3-bit MLX build is 194 GB. Consumer GPUs are out of range. This is workstation-or-datacenter territory.
- Training data cutoff and knowledge recency. As a Chinese-lab model, training data recency and Western-world knowledge may trail the Llama/Claude ecosystem for recent events outside China.
Real-world performance
- vs Qwen 3 235B-A22B: Hunyuan Large wins Chinese language performance; Qwen 3 wins English benchmarks, licensing (Apache 2.0), and ecosystem breadth. Pick by dominant language in workload.
- vs DeepSeek V4 Pro: Both Chinese labs, both frontier MoE. V4 Pro is MIT, while Hunyuan's license excludes the EU and requires a separate license above 100 million MAU.
Should you run this locally?
Yes if your workload is Chinese-language-heavy (>50% of tokens), you have workstation-class hardware (Mac Studio M3 Ultra 256 GB with the 3-bit MLX build), and Tencent's license terms work for your deployment. For Chinese-language operators, Hunyuan Large is genuinely one of the strongest open-weight options.
No if your workload is predominantly English (pick DeepSeek V4 Pro or Qwen 3 235B-A22B), you need Apache 2.0 licensing (pick Qwen), or you lack workstation-class hardware.
How it compares
- vs Qwen 3 235B-A22B: Hunyuan wins Chinese; Qwen wins English, licensing, ecosystem.
- vs DeepSeek V4 Pro: Hunyuan wins Chinese edge; DeepSeek wins all-around balance and ecosystem maturity.
- vs GLM-5: Both Chinese-frontier models. GLM has different architectural lineage (Tsinghua/Zhipu).
Run this yourself
- Datacenter: 8× H100 SXM (640 GB) at FP8 (389.55 GB) with Tencent's vLLM fork.
- Workstation: Mac Studio M3 Ultra (256 GB) with the 194 GB mlx-community 3-bit build. The realistic consumer-tier path.
- Cloud rental: Tencent Cloud TI has integrated Hunyuan-Large, or rent H100 clusters.
- Vendor: huggingface.co/tencent/Tencent-Hunyuan-Large
Overview
Tencent's frontier MoE. 389B total / 52B active. License permits commercial use with restrictions on companies above MAU thresholds.
How to run it
Hunyuan-Large is Tencent's 389B MoE model (52B active). Tencent ships two inference backends: its vLLM fork and TensorRT-LLM. Its README says BF16 requires 16 H20 GPUs, and it also publishes FP8 weights (389.55 GB). llama.cpp does not register this architecture (HunYuanForCausalLM) and Hugging Face lists no GGUF, so there is no Q4_K_M route. mlx-community publishes a 3-bit MLX build at 194 GB. Context: 256K on the pre-trained model, 128K on Instruct.
Hardware guidance
Minimum for the FP8 checkpoint: 5× A100 80GB (400 GB) for the 389.55 GB of weights alone. Tencent's BF16 deployment requires 16 H20 GPUs. Apple Silicon path: the mlx-community 3-bit build (194 GB) fits a 256 GB Mac Studio M3 Ultra. VRAM math: 389B total, 52B active. All experts must stay accessible, so memory follows the total, not the active count. RTX 4090 24GB and dual RTX A6000 (96 GB) are far below the weights.
What breaks first
- No llama.cpp path. llama.cpp does not register HunYuanForCausalLM and Hugging Face lists no GGUF. 2. Chinese-language bias. Hunyuan-Large is Tencent's model, and training data is Chinese-heavy. English quality is competitive but may show Chinese-culture bias in nuanced prompts. 3. AWQ on MoE. AWQ-INT4 quantization on MoE architectures can produce worse degradation on expert-routing stability versus dense models. Test routing correctness at Q4 before deploying.
Runtime recommendation
Common beginner mistakes
Mistake: Assuming "52B active" means it fits in 32 GB VRAM. Fix: All 389B experts must be accessible. The 52B is what's computed per token, not what's stored. The FP8 checkpoint is 389.55 GB. Mistake: Planning FP8 on too few GPUs. Fix: The 389.55 GB of weights alone need 5× A100 80GB (400 GB). Mistake: Ignoring license restrictions. Fix: Tencent's license does not apply in the EU, and services above 100 million monthly active users must request a license. Verify on huggingface.co/tencent/Tencent-Hunyuan-Large before production deployment.
Strengths
- Open-weight frontier MoE
- Strong on Chinese + English
Weaknesses
- Multi-machine cluster required
- Tier-restricted commercial license
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|
At Q4 Hunyuan Large 389B MoE needs about 346 GB of memory by our estimate, more than a 32 GB graphics card holds. Renting a GPU with enough memory (or several GPUs) by the hour is the practical way to try it before you plan hardware for it.
Referral links: RunLocalAI may earn a commission or account credits if you sign up and rent, at no extra cost to you. Hourly rates and which GPUs are listed change all the time, so check the provider's page before you start. How we make money.
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Hunyuan Large 389B MoE.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
Can I use Hunyuan Large 389B MoE commercially?
What's the context length of Hunyuan Large 389B MoE?
Source: huggingface.co/tencent/Tencent-Hunyuan-Large
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Hunyuan Large 389B MoE runs on your specific hardware before committing money.