Tencent HY4: Pricing, Vision, Coding Tests and Open Weights
Tencent HY4 hands-on: coding tests, new vision support, $2.50 output pricing versus DeepSeek V4.1 Flash, and whether open weights are coming.
For Engineering and IT
How good is Tencent HY4, and will it be open-weight?
Tencent HY4 is the full release of the model first shown as HY4 preview. It improves on the preview across Tencent's published benchmarks and adds image input for the first time. Output costs $2.50 per million tokens by the reviewer's account, which is not cheap next to rivals. The final HY4 weights are not public yet. The preview weights are, under Apache 2.0, so a release of the final model looks likely but is unconfirmed.
TL;DR
- HY4 shows a clear jump over the preview on the benchmarks where Tencent plots the preview as a dotted baseline.
- It supports vision perception for the first time, so it can take images as input, which matters for 3D and robotics-style tasks.
- Hands-on tests of a browser OS, a C++ skateboard game, a Godot wrestling game and a subway FPS came out acceptable to very good, with the skateboard game the standout.
- Output pricing of $2.50 per million tokens faces stiff competition, and DeepSeek V4.1 Flash is likely cheaper.
- The HY4 preview is open-weight (770B MoE, 49B active), but the final HY4 is not released as weights yet.
- Tencent's own materials admit a gap to frontier closed models and some known issues like overlong reasoning.
What changed from HY4 preview to HY4?
The preview launched about two months before the full release. Tencent's benchmark chart for HY4 marks the preview's scores with a dotted line, and HY4 sits above it on the benchmarks where that line appears. The reviewer who tested both called it a substantial capability leap.
The headline new feature is vision. Tencent says HY4 supports vision perception for the first time. The preview was text-only as far as its model card describes. Image input should help with front-end and 3D generation work, and with robotics tasks such as controlling a simulated arm.
Tencent also leans on front-end and game development in its benchmark presentation. And it openly shows a remaining gap between HY4, other open-weight models and the top closed frontier models. That candor is useful when you are deciding whether to trust the rest of the chart.
One more detail: the reviewer says Tencent trimmed some cybersecurity material from the training data. Patching and offensive topics stay in, but reverse engineering appears to be weakened. If you use models to poke at firmware on hardware you own, test that before you commit.
How did HY4 do in hands-on coding tests?
The reviewer ran most tests at the highest effort setting, using a free API key supplied by Tencent. That is a disclosure worth keeping in mind. The reviewer also notes they did not pay for the runs, so the cost shown next to each result is indicative only.
| Test | What it asked for | Result |
|---|---|---|
| Browser OS | A desktop-style OS in the browser with apps and two games, the second meant to be creative | Acceptable. Terminal, text editor, paint app, settings and a seed-driven world all worked. The GTA-style game had odd traffic and police behavior. |
| C++ skateboard game | Self-contained C++ game on an early 2000s New York block, with a replay on the C key | Strong. Sound, pedestrians, a fire hydrant and a cinematic replay all worked. Left and right controls were inverted. |
| Godot and Blender wrestling | A 1990s-style wrestling game built with Blender and Godot | Good. Distinct wrestlers, combos, finishers, taunts and generated textures. Sound effects were odd. |
| Subway FPS | Train-based waves, two weapons, zoom, ammo pickups | Solid. Some wave audio from a different game bled in, and the train handles did not move with the train. |
Browser OS
The generated "Nova OS" had a clock, sound toggle, dock, terminal, notes, a paint app with brushes, spray and symmetry modes, and PNG export. A right-click menu worked, and changing the seed altered the games. Nothing was mind-blowing, nothing was broken. The reviewer shortened the prompt for this round and told the model to make its second game creative. The result was a color-matching gate game, which was more original than the usual space shooter.
C++ skateboard game
This was the best result of the set. The reviewer was surprised enough to suspect the task might be benchmarked, and plans an unseen C++ test to check. That caution is reasonable. One good run on a familiar prompt does not show general C++ skill without dependencies.
Wrestling and subway FPS
In the wrestling game, running into the ropes knocked the wrestler out, a quirk the reviewer had not seen before. The model also created textures through the Blender pipeline. In the FPS, enemy models were simple but the game was feature-packed. The audio crossover from another generated game was the main flaw.
Does HY4 really handle vision and Blender tasks?
Vision is confirmed as a feature. The sources I have do not show a finished vision-to-Blender or robot-arm test, so I won't claim results there. The reviewer said they were eager to run the robot arm test but that their workspace was under renovation. Treat image-driven 3D and robotics performance as untested until more results appear.
How much does HY4 cost?
The reviewer cites $2.50 per million output tokens. I don't have a confirmed input price or cache pricing from a primary source, so check Tencent's official pricing page before budgeting.
The reviewer's verdict is that this is not cheap. Closed frontier labs and open-weight providers both offer strong competition at that level. DeepSeek V4.1 Flash is named as an extremely strong model that should come in cheaper, though the reviewer hedges with "I do believe" rather than quoting a price.
Also factor in reasoning length. Tencent's own preview notes say the model can spend longer than needed on complex tasks and over-verify its work. High reasoning effort burns output tokens, and output tokens are what you pay $2.50 for. The preview's documentation describes a no_think setting for direct answers, and the final API may offer something similar.
Will HY4 be open-weight?
Not yet. The reviewer states HY4 is not currently open-weight, while HY4 preview is. The preview and an FP8 version are on Hugging Face, ModelScope, GitCode and CNB under Apache 2.0. According to Tencent's Hy4 preview model card, the company shipped Hy4 preview early, as it did with Hy3 preview, and the reviewer expects HY4 weights "in the near future." That is an inference. Tencent has not committed to a date in the material I reviewed.
How big is HY4?
Tencent has not given a size for the final model in the sources I have. The preview is the best guide.
| Property | HY4 preview |
|---|---|
| Architecture | Mixture-of-Experts |
| Total parameters | 770B |
| Active per token | 49B |
| Layers | 78 (first dense, 77 MoE) |
| Experts | 256 routed plus 1 shared, top-8 routed per token |
| Context length | 1M tokens |
| Attention | Gated DeepSeek Sparse Attention |
| Extras | One native multi-token-prediction layer for speculative decoding |
A vision component would add some parameters to the final model, so expect a size near the preview's, not identical. For self-hosting the preview, Tencent's vLLM example serves the FP8 build with tensor parallelism of 8, which implies an eight-GPU node. Plan on that class of hardware for HY4 as well.
How does HY4 compare to rivals?
Tencent's preview model card reports a blind evaluation: 163 internal experts rated outputs on 203 engineering tasks. The preview edged out GLM 5.3 (2.99 vs. 2.92 average, 46.8% wins, 12.8% ties, 40.4% losses) and Kimi K3 (2.99 vs. 2.94, 51.2% wins, 7.9% ties, 40.9% losses). These are the vendor's own numbers on Tencent-selected tasks from Tencent staff, so treat them as a signal, not a verdict. They also cover the preview, not the final HY4.
Frequently Asked Questions
Is Tencent HY4 open source?
The final HY4 is not open-weight at the time of the review. HY4 preview is, released under the Apache 2.0 license with an FP8 variant. The reviewer expects the final weights to follow, but no date is confirmed.
Does HY4 support images?
Yes. Tencent says HY4 supports vision perception for the first time, so it can take images as input. The preview did not have this described in its model card.
How much does the HY4 API cost?
The reviewer cites $2.50 per million output tokens. I could not confirm input pricing from a primary source. The reviewer considers the price not cheap, and thinks DeepSeek V4.1 Flash costs less.
Is HY4 better than the preview?
Tencent's benchmark chart shows HY4 above the preview's dotted baseline on the benchmarks where it is plotted, and the reviewer saw a substantial leap in practice. Exact scores are in Tencent's benchmark image, which I could not reproduce here.
Can I run HY4 locally?
Not yet, since the final weights are unreleased. The preview is a 770B MoE with 49B active parameters, and Tencent's own serving example uses eight GPUs with the FP8 build. That is a server-class setup, not a desktop one.
Topics
Useful for
How we source, write, and correct our articles. Editorial standards
