What Do You Need to Run Hunyuan Locally? VRAM by Version
Hunyuan currently has 2 versions you can deploy locally. The smallest, Hy3, needs 186GB of VRAM at 4-bit; the largest, Hy4 preview, is an MoE model and needs 19GB of VRAM plus about 452GB of RAM with MoE offloading. The cheapest device that gets one running is the 8× Tesla V100 32GB, at about $7,160.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| Hy4 previewMoE | Aug 28, 2026 | 770B | 49B | 1M | Apache-2.0 |
| Hy3MoE | Jul 6, 2026 | 295B | 21B | 256K | Apache-2.0 |
About Hunyuan
Hunyuan (shortened to Hy in international markets since 2026) is Tencent's large model family, developed by the Tencent Hunyuan team (Tencent was founded in 1998 and is headquartered in Shenzhen). Hunyuan debuted in September 2023 and began open-sourcing weights in 2024. Its main language models are all MoE, with a focus on agents, coding and office work.
Hunyuan-Large, open-sourced in November 2024, has 389 billion parameters with 52 billion active per token, and Tencent called it the largest open-source Transformer MoE at the time. Hunyuan-A13B, from June 2025, can switch between fast and slow thinking. In early 2026 Tencent rebuilt its training infrastructure. Hy3 preview, open-sourced in April, was the first model after the rebuild, and the full Hy3 release in July has 295 billion parameters with 21 billion active per token. Hy4 preview, in August, scaled up to 770 billion parameters with 49 billion active per token, using a gated version of DeepSeek Sparse Attention and a 1-million-token context. The license also went through one major change: Hy3 preview and earlier models use Tencent's own license, which does not apply in the EU and some other regions, while Hy3 and Hy4 preview moved to Apache 2.0, which allows commercial use.
- VendorTencent
- Versions2
- Parameters295B – 770B
- Latest versionHy4 preview
- License
- Apache-2.0
- Official site
- hy.tencent.ai
- Hugging Face org
- tencent
How much VRAM do Hunyuan models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a Hunyuan version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Hy4 preview770B (49B active) | 307.2est. | 311.2 | 320.1 | 331.9 | 402.5 |
| Hy3295B (21B active) | 107.2GGUF | 119 | 151.2 | 194.2 | — |
How much VRAM do Hunyuan models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a Hunyuan version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Hy4 preview770B (49B active) | 467.3GGUF | 471.3 | 480.2 | 491.9 | 562.6 |
| Hy3295B (21B active) | 182.2GGUF | 194 | 226.2 | 269.2 | — |
How much VRAM do Hunyuan models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a Hunyuan version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Hy4 preview770B (49B active) | 857est. | 861 | 869.9 | 881.6 | 952.3 |
| Hy3295B (21B active) | 317.7GGUF | 329.5 | 361.7 | 404.7 | — |
FAQ
Is Hunyuan open source and free for commercial use?
Yes. The versions listed on this site, Hy3 and Hy4 preview, are both licensed under Apache 2.0, so you can use them commercially, modify them and redistribute them, as long as you include the license, keep the copyright notices and mark any files you changed. The claim you may find online that “Hunyuan can't be used in the EU” is about older models. Hunyuan-A13B, for example, uses the Tencent Hunyuan Community License, which states that it does not apply in the EU, the UK or South Korea. The licenses for Hy3 and Hy4 preview contain no such clause.
Can you run Hunyuan with Ollama?
Not at the moment. Ollama's official library has neither Hy3 nor Hy4 preview, not even as cloud versions. Hy3 runs in llama.cpp with bartowski's GGUFs; mainline has supported it since July 2026, including MTP speculative decoding. Support for Hy4 preview only landed in mainline in September 2026, but the existing AngelSlim GGUFs still require a patched build of llama.cpp. Both models have official vLLM recipes, so running the official weights on vLLM is the safer route.
