AI model · Tencent
Tencent Hy4, Explained: Specs, Price, Weights
Tencent Hy4 is the Hunyuan team's flagship language model: the open preview is a 770B-parameter mixture-of-experts model with a 1M-token context, and the finished API model costs about $0.83 per million input tokens. Synexa does not serve it, but it does run Tencent's Hunyuan3D image-to-3D models from $0.025 a request.
- 770B total, 49B active
- 1M-token context
- Preview weights: Apache 2.0
- Synexa: Hunyuan3D from $0.025
What is Tencent Hy4?
Tencent Hy4 is the fourth-generation large language model from Tencent's Hy team, the group behind the Hunyuan models. Its open weights live under the Tencent-Hunyuan organization and carry a hunyuan tag, which is why people search for it as Hunyuan 4 or hunyuan4 as often as by its new name.
It came out in two steps. The Hy4 preview shipped as open weights in late August 2026: a 770B-parameter mixture-of-experts (MoE) model that activates 49B parameters per token, released under the Apache 2.0 license. About two months later, in early November 2026, Tencent released the finished Hy4. The final model adds image understanding and computer use, meaning it can read photos and diagrams and operate desktop software on its own.
The finished Tencent Hy4 is API-only for now. Tencent says its weights will follow at a date it has not set, together with a safety report.
Tencent Hy4 versions and release dates
The Hy line has moved fast: three releases in about four months.
Hy3
- Released
- July 2026
- What it is
- 295B-parameter model, Apache 2.0
- Weights
- Open
Hy4 preview
- Released
- Late August 2026
- What it is
- 770B MoE, 49B active, 1M context, text only
- Weights
- Open, Apache 2.0 (plus an FP8 build)
Hy4
- Released
- Early November 2026
- What it is
- Adds image input and computer use, trained on more data
- Weights
- API only; weights planned
Hy4 flash
- Released
- Announced, not released
- What it is
- A faster, cheaper variant Tencent says is next
- Weights
- Not listed
| Model | Released | What it is | Weights |
|---|---|---|---|
| Hy3 | July 2026 | 295B-parameter model, Apache 2.0 | Open |
| Hy4 preview | Late August 2026 | 770B MoE, 49B active, 1M context, text only | Open, Apache 2.0 (plus an FP8 build) |
| Hy4 | Early November 2026 | Adds image input and computer use, trained on more data | API only; weights planned |
| Hy4 flash | Announced, not released | A faster, cheaper variant Tencent says is next | Not listed |
Dates are Tencent's public release dates. Tencent has not published a separate parameter count for the final Hy4.
Tencent Hy4 specs: architecture, context and license
Architecture figures are for the open Hy4 preview backbone, excluding the multi-token prediction layer. As of November 2026.
- Maker
- Tencent Hy team (Hunyuan)
- Architecture
- Mixture-of-Experts, 78 layers (first dense, 77 MoE)
- Parameters
- 770B total, 49B active per token
- Experts
- 256 routed plus 1 shared; 8 routed experts active per token
- Attention
- Gated DeepSeek Sparse Attention with cross-layer index reuse
- Context length
- 1M tokens
- Vocabulary
- 120,832 tokens
- Input / output (final Hy4)
- Text and images in, text out
- Reasoning
- Defaults to high; a
no_thinksetting gives direct answers - License (preview)
- Apache 2.0
- Builds
- Hy4 preview and Hy4 preview-FP8
How much does Tencent Hy4 cost?
Tencent Hy4 costs about $0.83 per million input tokens and $2.50 per million output tokens, the same price as the preview. Approximate dollar prices as of November 2026.
Input, per million tokens
- Tencent Hy4
- about $0.83
- Notes
- Same as the Hy4 preview
Output, per million tokens
- Tencent Hy4
- about $2.50
- Notes
- Same as the Hy4 preview
Example: 100K in + 10K out
- Tencent Hy4
- about $0.11
- Notes
- 0.1 × $0.83 + 0.01 × $2.50
Self-hosting the preview
- Tencent Hy4
- $0 license fee
- Notes
- Apache 2.0; you pay for the GPUs
Consumer chat
- Tencent Hy4
- Yuanbao app
- Notes
- Price not listed here
| Item | Tencent Hy4 | Notes |
|---|---|---|
| Input, per million tokens | about $0.83 | Same as the Hy4 preview |
| Output, per million tokens | about $2.50 | Same as the Hy4 preview |
| Example: 100K in + 10K out | about $0.11 | 0.1 × $0.83 + 0.01 × $2.50 |
| Self-hosting the preview | $0 license fee | Apache 2.0; you pay for the GPUs |
| Consumer chat | Yuanbao app | Price not listed here |
Dollar figures are approximate and may differ by platform. For scale, the same 100K-in, 10K-out call costs $1.20 on Claude Fable 5.2 at list price.
What is Hunyuan 4 good at?
Tencent built the training data with in-house engineers, game developers, finance analysts and security experts, so its strengths follow that work.
Long software tasks
Game development
Office and finance
Image understanding
Computer use
Research help
Where Tencent Hy4 falls short
Tencent is unusually direct about the gaps. It says the finished Hy4 still trails the best commercial models by a wide margin on advanced math, longer coding projects and complex image analysis. The preview shipped with two known habits: it reasons longer than needed on hard tasks, and it tends to over-verify its own work, both of which cost output tokens.
There are practical limits too. The final model's weights are not out yet, so you can't self-host the version with vision. And the open preview is big: Tencent's own serving recipes run the FP8 build across 8 GPUs with tensor parallelism. If you need a small, cheap model for chat or classification, wait for the announced Hy4 flash or pick a smaller model. If you are choosing a multi-model API platform rather than one model, our WaveSpeed AI alternative page compares prices model by model.
How to use Tencent Hy4: API or self-hosted
You have three ways in, depending on whether you want the final model or open weights.
Call the API
Self-host the preview
Chat with it
temperature 0.9 and top_p 1.0, and you can pass reasoning_effort set to no_think for direct answers without a long chain of thought.Tips for running hunyuan4 in production
What to plan for before you put it behind a product.
- Use the FP8 build to cut GPU memory when self-hosting
- Turn reasoning down to
no_thinkfor simple requests to save output tokens - Keep web search on for factual questions
- Re-run your evals when the final weights ship
- Expect the open preview to read images; that arrives with the final release
- Budget for one GPU; Tencent's recipes use 8
- Rely on it for hard math or long coding projects without your own evals
- Assume the API price includes the planned flash variant
Does Synexa serve Tencent Hy4?
No. Synexa does not serve Tencent Hy4 or any other chat language model. What it does serve is Tencent's 3D side: the Hunyuan3D image-to-3D models, through the same pay-per-request API as the rest of the catalog.
- Hunyuan3D 3.1: $0.16 a request, Tencent's newer image-to-3D model
- Hunyuan3D 2: $0.025 a request, the open-weights generation
So a common setup is to let a language model such as Hy4 write a game prop list or a product description, then send the reference images to Synexa to get 3D models back. See every Tencent model on Synexa under Tencent models, and start with $0.10 of free credit. Your first call takes a few lines of Python or Node with the quickstart.
- Hunyuan3D 3.1 at $0.16 a request
- Hunyuan3D 2 at $0.025 a request
- One API key for image, video, audio and 3D
Related pages
Other new models and API comparisons.
Frequently asked questions
What is Tencent Hy4?
Tencent Hy4 is the fourth-generation flagship language model from Tencent's Hy (Hunyuan) team. The preview, released as open weights in late August 2026, is a 770B-parameter mixture-of-experts model with 49B active parameters and a 1M-token context. The finished model followed in early November 2026 and added image understanding and the ability to operate software.
Is Hy4 free?
Partly. The Hy4 preview weights are free to download and use under the Apache 2.0 license, but running a 770B model needs a multi-GPU server. The finished model is a paid API at about $0.83 per million input tokens and $2.50 per million output tokens as of November 2026. Tencent's Yuanbao app also offers chat with it.
Is Hunyuan 4 the same as Hy4?
Yes. Hy is the name Tencent's Hunyuan team now uses for its language models, so Hy4 is the model people call Hunyuan 4 or hunyuan4. The open weights are published under the Tencent-Hunyuan organization and tagged hunyuan. Don't confuse it with Hunyuan3D, a separate Tencent model family that turns images into 3D models.
Can I download the Tencent Hy4 weights?
You can download the preview: Hy4 preview and an FP8-quantized build are on Hugging Face and ModelScope under Apache 2.0. The finished Hy4, with image input and computer use, is API-only for now. Tencent says it will release those weights later, with a detailed safety report, but has not given a date.
What hardware do I need to run Hy4 preview?
Plan for a multi-GPU server. Tencent's reference vLLM and SGLang commands serve the FP8 build with tensor parallelism across 8 GPUs, plus speculative decoding through the model's multi-token prediction layer. Only 49B of the 770B parameters are active per token, which keeps generation fast, but all weights still have to fit in GPU memory.
How does Tencent Hy4 compare with Claude Fable 5.2?
They aim at different budgets. A call with 100K tokens in and 10K out costs about $0.11 on Tencent Hy4 and $1.20 on Claude Fable 5.2 at list prices. Tencent itself says Hy4 trails the best commercial models on hard math, long coding projects and complex image analysis, while Fable 5.2 is Anthropic's top model for exactly those long, hard tasks.
Does Synexa offer Tencent models?
Yes, on the 3D side. Synexa serves Tencent's Hunyuan3D 3.1 at $0.16 a request and Hunyuan3D 2 at $0.025 a request for turning images into 3D models. It does not serve Tencent Hy4 or any other chat language model. New accounts get $0.10 of free credit and pay per request, with no subscription.
Run Tencent's 3D models on Synexa
Tencent Hy4 handles text and code; Synexa runs Tencent's Hunyuan3D models from $0.025 a request, with image, video and audio models on the same key. Start with $0.10 free credit.