HUNYUAN · TENCENT'S OPEN MODEL FAMILY
Hunyuan is now Hy3 and Hy4 — and that is why you cannot find it
Tencent open-sourced Hy3 on 6 July 2026 under a plain Apache 2.0 licence, then previewed Hy4 on 28 August. Most English explainers still call the family Hunyuan and stop there, which is why the naming is the first thing worth knowing.
Direct answer
What Hunyuan actually is, in one paragraph
Hunyuan is Tencent's model family, and its current generation ships as Hy3 and Hy4 rather than under a Hunyuan version number. Hy3 is a mixture-of-experts model with 295B total parameters and 21B active — a 14-to-1 sparsity ratio that is the whole point, because it aims at the quality of models several times its size while only paying to run a fraction of them. It handles 256K tokens of context and was released under Apache 2.0.
Comparison basis
Sources
Release dates, model sizes, context length and licence come from Tencent's own releases. VUST has no relationship with Tencent and runs none of these models.
Unless a separate method is stated above, this page does not claim a head-to-head product-quality benchmark.
The Hunyuan lineup
Hy3
Released
6 July 2026
Shape
295B total · 21B active · 256K context · Apache 2.0
What it is for
General reasoning and agentic work at open-weight cost
Hy4 preview
Released
28 August 2026
Shape
770B total · 49B active
What it is for
Coding agents, tool use, longer workflows
Hunyuan A13B Instruct
Released
Earlier generation
Shape
Mixture of experts
What it is for
The compact entry, still widely served
Hy-MT2 translation models
Released
Same family
Shape
Three sizes: 1.8B, 7B, 30B-A3B
What it is for
Dedicated translation rather than general chat
Where you can actually run it, and what VUST runs instead
The weights are open, so Hy3 is served by independent hosts rather than by Tencent alone, at a small fraction of a frontier flagship's token price — that gap, not any single number, is what the open-weight tier is for. VUST does not carry Tencent's models. From the same tier, selectable by name in @vustbot: DeepSeek V4.1 Flash, GLM-5.3, Kimi K3, DeepSeek V4 Pro, MiniMax M3, Qwen 3.7 Plus, Qwen 3.7 Max.
The lineup in short
Tencent's Hunyuan family now ships as Hy3, open-sourced 6 July 2026 under Apache 2.0 with 295B total and 21B active parameters and a 256K context window, and Hy4 preview, released 28 August 2026 with 770B total and 49B active parameters. Because the weights are open, the models are served by multiple independent hosts at open-weight prices.See the open-weight models in @vustbot
Frequently asked questions
Why can I not find a model called Hunyuan 3?
Because Tencent stopped putting the family name in the version. The current models are Hy3 and Hy4; Hunyuan is the family they belong to. Most English write-ups have not caught up, which is why the query returns pages that never mention the rename.
Is Hunyuan open source?
Hy3 was released with open weights under Apache 2.0 in July 2026, and that licence allows commercial use. Open weights are not the same as an open dataset or open training code.
Can I use Hunyuan in VUST?
No — VUST runs none of Tencent's models. For an open-weight model in Telegram today, @vustbot runs DeepSeek V4.1 Flash, GLM-5.3, Kimi K3, DeepSeek V4 Pro, MiniMax M3, Qwen 3.7 Plus, Qwen 3.7 Max, each selectable by name.
What does the 295B/21B split mean in practice?
It is a mixture-of-experts design: the model holds 295 billion parameters but activates about 21 billion for any given token. You get the breadth of a very large model while paying compute for a small one, which is why open-weight hosts can serve it cheaply.
Is Hy4 better than Hy3?
Hy4 is larger and aimed at coding agents and tool use, and it is still a preview. Larger and newer is not automatically better for your task, and a preview is a different reliability proposition from a stable release.