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

A frontier-scale model, given away under a permissive license

On August 28, 2026, Tencent's Hunyuan team released Hy4 preview and published its weights the same day. The new flagship is a Mixture-of-Experts model with 770 billion total parameters, 49 billion of which are active per token, and a context window of more than one million tokens.34 Tencent puts the release out under the Apache License 2.0.1 That license matters about as much as the parameter count. One analysis of the model's local-deployment economics called it the most permissive license of any model this size, with no usage clauses attached.19

The weights come in two versions: a full instruct model and an FP8-quantized build. Both are hosted on Hugging Face, ModelScope, GitCode and CNB.1 Tencent's announcement post summed up the pitch in a few words: built for productivity, positioned at the open-source frontier, sold at a steady low price, with an invitation for users to report what breaks.10

Tencent wraps all of this in the idea of "deep model-product co-design." Most of the coverage repeats the slogan without asking much about it. My read is that the phrase points to something real about how Tencent now builds models. It also explains why releasing the weights works in Tencent's favor rather than against it.

What actually shipped

The architecture details on the GitHub repository are unusually complete. The backbone has 78 layers. The first uses a standard dense feed-forward block, and each of the other 77 holds 256 routed experts plus one shared expert. Every token goes to the top eight routed experts and the shared one.1 A separate multi-token prediction layer, with 10 billion parameters and 0.7 billion active, is included for speculative decoding.1

Tencent credits rival labs openly. The attention module uses Gated DeepSeek Sparse Attention with an IndexCache mechanism that lets layers reuse sparse indices, and the repository names DeepSeek and GLM as its inspirations.1 Western commentary picked up on this. One outlet described it as a rare acknowledgment of cross-pollination between competing Chinese labs, and read it as evidence that those labs are converging on the same answer to the cost of very long context.15 Open weights are what make that kind of borrowing possible. Hy4 builds on published techniques from DeepSeek and GLM, and its own weights now go back into the same shared pool.

The release also comes ready to run. Tencent provides prebuilt Docker images for vLLM and SGLang that expose an OpenAI-compatible API, and the reference setups use eight-way tensor parallelism.1 Developers can start serving the model on day one instead of waiting for the community to write inference support.

The co-design angle, decoded

Tencent's case for co-design has two parts. First, the training data was built with in-house experts in software engineering, gaming, finance and security. Second, the model was developed alongside products such as WorkBuddy and CodeBuddy.5 The company describes this as a loop: product needs shape the model, and real-world performance and user feedback shape the next version.5

The main evidence Tencent offers is a blind evaluation run inside WorkBuddy. In it, 163 internal experts rated outputs on 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, against 2.94 for Moonshot's Kimi K3 and 2.92 for Zhipu's GLM 5.3.36 Every outlet reports these numbers the same way. They disagree on how much the numbers prove. Some regional finance coverage treats the result as confirmation of Hy4's public benchmark gains.3 More skeptical analysts point out that the margin is 0.07 points on a four-point scale, and that Tencent designed, ran and graded the test itself.11 Another write-up notes that internal blind tests tend to reflect the evaluator's own mix of tasks.6

I side with the skeptics on what the test proves, but I think it says something else useful. A model co-designed with WorkBuddy and tested on WorkBuddy tasks should do well on WorkBuddy tasks. That is close to circular. It does show where Tencent's priorities are: agentic coding and office workflows inside its own products, more than leaderboard rankings. The public numbers point the same way. Tencent reports 85.4 on Terminal Bench 2.1, and a jump on DeepSWE from 28.0 for Hy3 to 64.3.3 As independent trackers note, all of these figures are still vendor-reported until someone reproduces them.17

Why open-sourcing helps Tencent

Why give away a 770-billion-parameter model that you also sell through an API? The pricing gives part of the answer. Tencent charges $0.834 per million input tokens and $2.501 per million output tokens, with cached tokens at $0.042 per million.5 Chinese-language coverage gives the equivalent as 6 yuan and 18 yuan.3 The model is also available through Tencent Cloud TokenHub and OpenRouter, and inside Yuanbao, ima, WorkBuddy and CodeBuddy. The two coding tools are free for two weeks after launch.24

The open weights don't really compete with this business, because few people can run them. One hardware analysis estimates that even a 4-bit quantization produces a file of about 435 GiB, which puts the model in 512GB-machine territory and rules out almost all hobbyist hardware.19 In practice, the people who can self-host are enterprises and research groups. Commentary on the release identifies Chinese firms with compliance-sensitive workloads, in finance, healthcare and government-adjacent work, as natural adopters, because they cannot send data to foreign APIs.6

The larger strategic case appears in analysis of the Chinese open-model race. US export controls make compute scarce. A closed, API-only model has to win on inference infrastructure, which is where that scarcity hurts most. An open-weight model can instead win by spreading: developers worldwide fine-tune it and build on it.11 Open weights also let a lab show frontier capability to developers who would hesitate to send sensitive workloads through a Chinese-hosted API.11 Seen this way, Apache 2.0 is Tencent's distribution strategy, not a gesture of generosity.

The timing and cadence

Hy4 preview arrived during a crowded week. Two days earlier, Alibaba shipped Qwen3.8-Flash-Next and Z.AI released GLM-5.3-Flash.11 Another write-up adds that ByteDance's Doubao Work launch fell in the same stretch, and argues that the major Chinese labs have adopted continuous open releases as their default competitive posture.6

Tencent's own pace is fast. Since rebuilding its infrastructure in February 2026, Hunyuan has shipped a major version roughly every two months, with a preview first and a full release later.2 The sources disagree on some details of the history. One account dates Hy3 preview to July 2026.11 A reference summary places Hy3 preview in April and the finished Hy3 on July 6.12 That summary also notes that Hy3 preview came out under a more restrictive Tencent Hy Community License, while the finished Hy3 and now Hy4 preview use Apache 2.0.12 If that is accurate, Tencent's licensing has loosened over the year, and this is the most important open-source detail of the release.

The jump in scale is large. Hy3 had 295 billion total parameters, 21 billion active, and a 256,000-token context window. Hy4 preview more than doubles the first two and roughly quadruples the context.17 One early write-up hedged on the license and pointed readers to the model card.14 The repository itself is clear that it is Apache 2.0.1

The caveats Tencent states itself

Tencent describes the model as unfinished. The repository lists known problems: it spends longer than needed reasoning through complex tasks and tends to over-verify its own work.1 Coverage also notes that Hy4 preview has no vision support.3 The company's stated reason for shipping a preview is that early real-world feedback made Hy3 much better, and it hopes for the same with Hy4.1

The most notable claim in the release is that Hy4 preview helped build itself. Tencent says the model took part in automatically optimizing training methods, data strategies, evaluation frameworks and low-level operators. It also tuned its own inference stack through operator fusion and communication improvements, which Tencent says raised end-to-end throughput by 31.8% over baseline.5 Like the benchmarks, this has not been independently verified.11

Open-sourcing does not mean the same thing across all of Tencent's model lines. When Hy Image 3.5 preview launched in September, it was available only through the API. The most-read replies under the announcement asked whether open weights were coming, and Tencent had not answered when one outlet checked.16 Hy4's text weights are open now. Whether the rest of the Hy lineup follows is still undecided.

The bottom line

Hy4 preview is a credible frontier-scale model with a permissive license and good deployment tooling on day one. Its claims to be best in class rest mostly on Tencent's own testing. The co-design story explains why it is strong at agentic office and coding work. The open weights extend that work beyond Tencent's own products. The next test comes as independent leaderboards and enterprise deployments arrive, and as the full Hy4 release replaces this preview.115