Hunyuan Hy4-preview
Tencent · Text generation · MoE 770B (A49B) · 1000k context · Released 28 August 2026
Hy4-preview is the preview release of Tencent's Hunyuan 4 flagship, published on 28 August 2026 across Hugging Face, ModelScope, GitCode and CNB. It is a sparse mixture-of-experts with 256 routed experts and one shared expert across 78 layers, routing each token to eight experts, and it carries a 1-million-token context and a reasoning mode that defaults to deep chain-of-thought and can be turned off. It is text-only. Tencent shipped it in both BF16 and FP8. The important licence point is that, read at source, it is standard Apache 2.0 with no geographic restriction, in contrast to Hunyuan media models such as HunyuanVideo whose licence excludes the EU, UK and South Korea. At 770 billion parameters it is a GPU-server model, not a consumer-card one.
Strengths
- Genuine Apache 2.0, worldwide, with no territory restriction and no revenue gate
- Very long 1-million-token context window
- Sparse: only about 49 billion of 770 billion parameters compute per token
- Released in both BF16 and FP8, with an optional reasoning mode
Weaknesses
- Server-class, needing a multi-GPU server rather than a workstation card
- Text-only, with no multimodal input or output
- Headline capability claims are Tencent's own, with independent testing still limited
- Reasoning mode is on by default, adding latency and token use unless turned down
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| Q4 (approx) | ~385GB | A community 4-bit quantisation, roughly 385GB (for the 770B total; verify against a specific GGUF). This fits a 512GB unified-memory machine, so the model is not datacentre-only. Figure is approximate. |
| FP8 (as released) | ~770GB | One of the released formats, roughly 770GB, close to one byte per parameter. Multi-GPU server territory, with actual needs rising with context and concurrency. Figure is approximate. |
| BF16 (as released) | ~1500GB | The full-precision release, roughly 1.5TB. Figures are approximate. |
What you'd need to run this
Licence
Apache 2.0 — read the licence
Hunyuan Hy4-preview: common questions
- What hardware do I need to run Hunyuan Hy4-preview?
- At its most compressed (Q4 (approx)) it needs roughly 385GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is Hunyuan Hy4-preview free for commercial use?
- Yes. Hunyuan Hy4-preview is licensed under Apache 2.0, which permits commercial use with no meaningful conditions.
- Can I run Hunyuan Hy4-preview on Apple Silicon?
- It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
- What is Hunyuan Hy4-preview's context window?
- Hunyuan Hy4-preview has a context window of 1,000,000 tokens, about 1000k.
Availability
Recommended for
- Organisations with cluster-class infrastructure wanting a permissively licensed flagship
- Server-class self-hosting of coding and reasoning work with data kept in-house
- Long-context text tasks at scale
Related models
Related guides
Glossary
Catalogue entry last verified 28 September 2026. Specifications change; verify anything you are about to spend money on.