Every Hunyuan AI Model Explained and Compared (Nov, 2026)

The Hunyuan family

Tencent still calls the program Hunyuan, but new models ship as "Hy": Hy3, Hy4, Hy-MT2. The family has two halves that most write-ups blur.

One half is the hosted API on TokenHub. The other is a set of open-weight checkpoints on Hugging Face, ModelScope and GitCode, and the two halves do not line up one to one.

Hy4 is on the API but its weights are not public yet, while Hy-SimulMT and Hy-Memory are open releases with no API row.

Around the text models sit image, video, 3D and world models, several of them older Hunyuan-branded releases that still matter.

Hy4

Hy4 adds two things no earlier Hy flagship had: image understanding and computer use.

TokenHub describes it as the official release that improves on Hy4 preview for agent and long-running work, and adds visual understanding and computer-use agent abilities. Caixin reported the launch on November 4, 2026.

The API takes up to 960K tokens of input inside its 1M window and returns up to 64K. It supports deep reasoning with preserved thinking, structured output, function calling, caching and image input.

As of November 12, Tencent had not posted Hy4 weights on Hugging Face, so the open checkpoint for this generation is still Hy4 preview.

Hy4 preview

Tencent ran a blind test before shipping Hy4 preview: 163 of its own experts rated outputs on 203 engineering tasks. Hy4 preview averaged 2.99 against 2.92 for GLM 5.3 and 2.94 for Kimi K3, according to the Hy4 preview model card.

It is a Mixture-of-Experts model with 770B total parameters, 49B active per token, 256 routed experts and a 1M token context, and it borrows its sparse attention from DeepSeek.

The weights arrived on August 28, 2026 under Apache 2.0, with a ready vLLM image that serves the FP8 version across 8 GPUs.

Tencent lists two known problems: it reasons for longer than it needs to and it over-checks its own work.

If you compare it with the open models in our GLM model guide or Kimi model guide, keep in mind that the scores above are Tencent's own.

Hy3

The headline change in Hy3 is fewer made-up answers. On Tencent's internal tests its hallucination rate fell from 12.5% to 5.4%, and its multi-turn issue rate from 17.4% to 7.9%, according to the Hy3 model card.

Tencent gathered feedback from more than 50 of its own products after the April preview, then scaled up post-training.

In a blind test with 270 experts, Hy3 scored 2.67 out of 4 against 2.51 for GLM-5.1. The architecture is the same as the preview: 295B total parameters, 21B active, 192 experts with 8 active per token, and a 256K context.

Tencent recommends 8 GPUs with large memory, such as H20-3e, to serve it. The Beijing News reported the July 6 launch and that Hy3 already runs inside WorkBuddy, CodeBuddy, Yuanbao and ima.

Hy-MT2 translation models

Hy-MT2 is Tencent's translation line, and its smallest model squeezes into 440 MB. Tencent open-sourced three sizes on May 21, 2026: 1.8B, 7B and a 30B-A3B MoE, all covering 33 languages under Apache 2.0, per the Hy-MT2 model card.

The 440 MB figure is the 1.8B model after 1.25-bit quantization, built for on-device use. On the API, Hy-MT2 comes in Lite, Plus and Pro tiers with an 8K context and 4K output. Tencent cut the Plus price on August 10, 2026.

Specialist models

Three smaller open releases fill narrow jobs. HunyuanOCR 1.5, about 1.1B parameters, parses documents, spots text in images and translates text inside pictures, with 4K image input and a 128K context.

Hy-SimulMT-30B-A3B, released October 26, 2026, interprets live speech between Chinese and English and clones the speaker's voice.

Hy-Memory-30B-A3B, with about 3B active parameters, replaces the general LLM inside memory frameworks such as Mem0 and MemOS and decides what an agent should remember. Hy-SimulMT and Hy-Memory are Apache 2.0.

HunyuanOCR uses the Tencent Hunyuan Community License.

Image, video, 3D and world models

Tencent calls HunyuanImage 3.0 the largest open-source image generation MoE model to date, with 64 experts, 80B total parameters and 13B active, according to its model card.

It generates images inside an autoregressive language model rather than a diffusion transformer, and the Instruct version added in January 2026 handles image editing and prompt rewriting.

On the API the image line has moved on. The hy-image-v3 endpoint goes offline on November 12, 2026, and Hy-Image 3.5 preview costs $0.024 an image at 1K or 2K and $0.032 at 4K.

HunyuanVideo 1.5 has 8.3B parameters, and its step-distilled 480p image-to-video model can render a clip within 75 seconds on a single RTX 4090. The hosted Hy-Video 1.5 costs $0.048 per second of 720p output.

The newest open video release, Hy-Nomas, builds a world one clip at a time from text prompts and camera moves, at 720 by 1280 and 24 frames per second, under Apache 2.0.

For 3D, HY-World 2.0 turns text, images or video into 3D meshes and Gaussian splats, first released in April 2026 under its own community license. The API runs HY-World 2.1 at $0.77 per panorama and $12.30 per walkable scene.

HY-3D asset models bill in credits at $0.02 each, and a HY-3D 3.0 or 3.1 request takes 15 to 60 credits, or $0.30 to $1.20.

Context windows and output limits

Hy4 reads about four times as much as Hy3 but writes half as much in one reply. Its 1M token window allows up to 960K tokens of input and 64K of output.

Hy3's 256K window allows 192K of input and 128K of output, which suits long code generation or a full report in a single call.

The Hy-MT2 tiers stop at 8K, with 4K in and 4K out, so long documents need to be split before translation.

The open checkpoints match the API. The Hy4 preview card lists a 1M context length and the Hy3 card lists 256K. Serving the full window yourself costs GPU memory, which is where most self-hosting budgets go wrong.

Which model should you use

If I were starting a new build on Hunyuan today, I would put Hy3 first. It costs a fifth of Hy4 on output, writes longer replies, and has open weights you can move in-house later.

Send work up to Hy4 when a task needs images, more than 192K tokens of input or computer use.

Measure before you escalate. Run a sample of real requests on Hy3, check the answers, and send only the failures to Hy4. While the Hy4 promotion lasts the gap is smaller, so test both now and plan your budget on the January list prices.

What this means for hiring

With Hunyuan, the engineering work sits in three places: routing between Hy3 and Hy4, setting reasoning effort per route, and serving a 295B MoE model on your own GPUs if data has to stay in-house.

Senior AI engineers who can do that cost $12,000 to $18,000 a month, all in, in the United States. Pre-vetted engineers in Vietnam or the Philippines with the same skills often cost $3,000 to $6,000.

Many of our clients run open models with an AI model deployment engineer or a wider AI engineering team from our network.

Build your Hunyuan stack with the right team

Hunyuan gives you a cheap hosted workhorse in Hy3, a multimodal flagship in Hy4, and Apache 2.0 weights for Hy3 and Hy4 preview. Picking the model is quick. The routing, serving and evaluation around it is engineering work.

We match you with pre-vetted senior engineers across Asia in about 24 hours, with no upfront cost and payroll handled.