Tencent Hy4 Preview Open-Sourced as 770B MoE Under Apache 2.0

Why Hy4 preview matters

Tencent's Hunyuan team, which now brands its models internationally as "Hy," released and open-sourced Hy4 preview on August 28, 2026. The model has 770 billion total parameters, 49 billion of them active per token, and a context window of more than one million tokens.1213 The weights went out under the Apache License 2.0, in a standard instruct version and an FP8-quantized version. Both are on Hugging Face, ModelScope, GitCode and CNB.11 Tencent switched to the shorter Hy name outside China in 2026 and still uses Hunyuan at home.4

The headline numbers drew most of the attention, but developers should look at how the release was put together. Tencent shipped open weights, ready-to-run inference images, an OpenAI-compatible API on its own cloud and on OpenRouter, and the model inside its own coding and productivity agents, all on the same day.1112 The company says it built the model through "deep model-product co-design" with CodeBuddy and WorkBuddy.1 I think that is the right way to read Hy4 preview: Tencent is arguing that an open model is only as good as the tools it was tuned against.

What actually shipped

The GitHub repository gives the architecture. There are 78 layers: the first uses a dense feed-forward block and the other 77 are mixture-of-experts layers, each with 256 routed experts and one shared expert. Every token goes to the top eight routed experts plus the shared one.11 A built-in multi-token-prediction layer, about 10 billion parameters with 0.7 billion active, handles speculative decoding.11 For attention, Tencent uses Gated DeepSeek Sparse Attention with an IndexCache that reuses sparse indices across layers. The residual path uses what Tencent calls identity Hyper-Connections, with four residual streams.11

Tencent openly says the design borrows from DeepSeek and Zhipu's GLM. Some coverage called that an unusual admission of cross-pollination between rival Chinese labs.5 It also shows the labs settling on similar ways to make million-token contexts affordable to run.5

The model is about twice the size of its predecessor. Hy3 had 295 billion total parameters, and active parameters have roughly doubled too.14 Tencent describes this as the biggest generation-over-generation jump it has measured, and says it puts the model at the open-source frontier.11

Practical for developers from day one

What stands out for engineers is how little setup is needed. The repository's quickstart talks to a local server through the standard OpenAI Python client. LMSYS publishes a prebuilt SGLang container, lmsysorg/sglang:hy4-preview, for both x86 and Arm. The sample launch command loads the FP8 weights with eight-way tensor parallelism, automatic reasoning and tool-call parsers, and speculative decoding turned on.11 Independent catalogs also list vLLM support with tool-call and reasoning parsers.17

The example uses eight GPUs, and that tells you who can self-host this. Hobbyists can't run a 770B MoE, even with only 49B parameters active. Teams with a multi-GPU node can, and the FP8 build is clearly aimed at fitting the model into that kind of hardware.11

The license may matter more than the hardware. Apache 2.0 lets anyone download, modify, fine-tune and use the model commercially.10 Tencent hasn't always been this permissive: the Hy3 preview came out under a custom Tencent Hy Community License, while Hy3 and Hy4 preview use Apache 2.0.4 That puts Tencent alongside the permissive licensing of DeepSeek (MIT) and Qwen (Apache 2.0). In this market a custom license is a real reason for teams to pick something else.1

If you'd rather not host it, the hosted API costs $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached tokens.12 Chinese-language coverage gives the price as 6 yuan for input and 18 yuan for output, roughly the same figures.14 On OpenRouter the model ID is tencent/hy4-preview.17 WorkBuddy and CodeBuddy are free for two weeks after launch, and free Hy3 access on both runs through September 30.18

Model-product co-design, and the evidence for it

Tencent explains Hy4 preview's gains by two things: training data built with internal experts in software engineering, gaming, finance and security, and close co-design with WorkBuddy.1218 The pitch is that improvements should show up in the daily work people do with Tencent's agents, not only on leaderboards.11

The evidence Tencent offers is a blind side-by-side test. 163 internal experts rated outputs on 203 engineering tasks, and Hy4 preview averaged 2.99 out of 4, against 2.94 for Moonshot's Kimi K3 and 2.92 for GLM 5.3.12 Chinese coverage adds that the test ran inside the WorkBuddy environment.14 That detail matters. A test in Tencent's own agent harness, with Tencent's own graders, is about the most favorable setting Hy4 could get.

Coverage splits here. Pandaily and TechBullion report the result mostly as Tencent framed it.1218 Other analysts are more cautious. One points out that a 0.07-point gap on a 4-point scale, designed and graded by the vendor, is a directional claim that needs third-party confirmation.1 Another warns that internal blind tests favor the evaluator's own mix of tasks.20 I side with the skeptics. The fair summary is that Tencent's own harness puts Hy4 roughly level with the best Chinese open models. It doesn't show Hy4 is better.

Public benchmarks Tencent reported tell a stronger story for coding agents, though they are also vendor-reported. Coverage cites 85.4 on Terminal Bench 2.1, a jump on DeepSWE from Hy3's 28.0 to 64.3, and 74.1 on Toolathlon-Verified.14 Several outlets stress that none of these figures had been independently reproduced at release.81

One outside signal has since appeared. On Arena's new Alignment Index, Hy4 preview scored 78.5, fifth worldwide and first among Chinese labs, with an unauthorized-action rate of 1.47%.3 The same report notes a false-completion rate of 13.24%, far behind the leader's 2.34%, and says the leaderboard's methods haven't been validated yet.3 For anyone building agents, that false-completion number deserves attention: it measures how often a model says it finished a task when it didn't.

Self-improvement claims

The boldest part of the launch is Tencent's claim that Hy4 preview helped build itself. According to the company, the model took part in automatically optimizing training methods, data strategies, evaluation systems and low-level operators. It proposed changes, ran experiments, and its code and logs fed into later rounds.12 Tencent also says the model profiled its own inference stack, reworked operator fusion and communication, and raised end-to-end throughput by 31.8% over a baseline.1318

Reporting agrees on what Tencent said and disagrees on how much to read into it. One write-up called it an early but real sign of where model development is going.20 Others treat it as an unverified vendor claim.1 I'd treat the throughput figure as the more concrete of the two. For developers, the useful question is whether those kernel improvements made it into the open SGLang and vLLM paths or stayed in Tencent's internal serving stack. The public materials don't answer that.

Known limitations

Tencent is unusually frank about the model's weaknesses. The repository says Hy4 preview often reasons longer than it needs to on complex tasks and tends to over-check its own work.11 Coverage adds that it has no vision input.14 Both matter in production. Over-long reasoning costs output tokens and latency, and at $2.501 per million output tokens those costs add up in long agent loops.12

The word "preview" means what it says. Since rebuilding its infrastructure in February, Hunyuan has put out a major iteration about every two months, releasing a preview first and a full version later, and it says more Hy4 models are coming soon.12 Sources disagree on the timeline: one dates Hy3 preview to July, while others put it in April with the finished Hy3 arriving July 6.14 The broader point is the same either way. This is Tencent's third major release in roughly six months.18

Where this leaves developers

The launch came in a busy week. Alibaba released Qwen3.8-Flash-Next and Z.AI released GLM-5.3-Flash two days earlier.1 That pace supports one analyst's advice not to tie a product to any single lab's model.1

In that crowded field, Hy4 preview has a few clear selling points: a permissive license, containers ready on release day, an OpenAI-compatible API, and a real agent product that serves as a live reference deployment. The co-design story is believable as a description of how Tencent built the model. As a guarantee that you'll get the same results in your own setup, it isn't proven yet. If your team has eight GPUs and a coding-agent workload, Hy4 preview is worth benchmarking. Test it on your own tasks before relying on Tencent's numbers.