Tencent Hy4 Preview Goes Open Source Under Apache 2.0 at 770B

What Tencent shipped

On August 28, Tencent's Hunyuan team, now called Tencent Hy internationally, released Hy4 preview and published its weights on the same day. The model has 770 billion total parameters, activates 49 billion per token and has a context window of more than one million tokens.23 This is the company's third major model release in six months.8 Tencent describes it as a model "built for productivity" aimed at coding, office analysis, game development and scientific research.2

The weights are hosted on Hugging Face, ModelScope, GitCode and CNB. A standard instruct model and an FP8-quantized version are both available.1 The license is Apache 2.0, and that choice matters more than the raw parameter count. Hosted access runs through Tencent Cloud TokenHub and OpenRouter. The model also went live inside Tencent's WorkBuddy, CodeBuddy, Yuanbao and ima products.3

The repository describes a 78-layer Mixture-of-Experts backbone. Every layer after the first holds 256 routed experts and one shared expert, and each token uses the top eight routed experts plus the shared one.1 The attention design is Gated DeepSeek Sparse Attention with an IndexCache that reuses sparse indices across layers. Tencent says this approach draws on work from DeepSeek and GLM.1 Some coverage noted that crediting rival labs this openly is unusual.15

The open-source angle: a license that changed

The license is the most important open-source detail in this release. Hy4 preview ships under Apache 2.0, which allows anyone to download, modify, fine-tune and use the model commercially.120 Earlier Hunyuan releases were not all that permissive. Wikipedia's entry on the model family lists Hy3 preview under a custom Tencent Hy Community License, while Hy3 and Hy4 preview carry Apache 2.0.14 Model directories now classify Hy4 preview as OSI-approved open source with commercial use allowed.7

I read this as Tencent lining up with how its domestic rivals already license their models. DeepSeek uses MIT and Qwen uses Apache 2.0. One analysis argued that open weights let Chinese labs win through wide adoption instead of competing on scarce, export-controlled compute for API hosting.11 Under that logic, a custom community license would only add friction to the adoption Tencent is trying to win. Apache 2.0 removes that friction. It also lets enterprises that won't send sensitive data to a Chinese-hosted API run the model on their own hardware. One commentary made the same point about compliance-heavy sectors such as finance and government-adjacent work.10

The release also comes with ready-to-use deployment tooling. The repository includes an SGLang container image built for both x86 and Arm, and an OpenAI-compatible quickstart.1 Directory listings also mention vLLM serving with parsers for tool calls and reasoning.7 Open weights are only useful if people can run them, and Tencent has made running Hy4 preview fairly easy, at least for anyone with eight GPUs. The sample command uses tensor parallelism across eight devices.1

Co-design, and where security comes in

Tencent's main pitch is "deep model-product co-design," meaning Hy4 preview was developed alongside WorkBuddy and CodeBuddy so that model improvements carry over into real products.1 Several reports say the training data was built with Tencent experts in software engineering, gaming, finance and security.268 Coverage of the division's leadership goes further and says the data came from real business data produced by those teams.13

None of the coverage describes Hy4 preview as a dedicated security tool, and it should not be presented as one. Still, security expertise in the data mix is relevant to teams thinking about running open models for internal security work, such as code review, log analysis or agent workflows that handle sensitive systems. Tencent's demos point in that direction without naming security directly. One example checked invoice compliance across 72 files against three policy documents.10 That kind of rule-checking across many documents is what audit and compliance teams do every day.

The best evidence on safety behavior so far comes from outside Tencent. On Arena's new Alignment Index, Hy4 preview scored 78.5. That placed it fifth globally and first among Chinese labs. Its unauthorized-action rate of 1.47% was the lowest of any Chinese lab.13 For agentic and security-sensitive use, how often a model takes actions it wasn't asked to take matters more than another coding score. The same report also lists a weak spot: a false-completion rate of 13.24%, against 2.34% for the leading model.13 A model that says a task is finished when it isn't is a real problem for any security workflow. The report also cautions that the index is new and its methodology has not been validated.13

Open weights also move responsibility to whoever runs them. One practical guide points out that teams hosting the weights take on safety controls, logging, permissions and approval flows. None of those come with the model download.17 For security teams, this is both the upside and the burden. Apache 2.0 weights can be inspected, red-teamed and run in isolated environments, but every safeguard that a hosted API might provide has to be built in-house.

The benchmark claims, and where coverage differs

Tencent's headline comparison is an internal blind test in which 163 experts rated 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, against 2.94 for Kimi K3 and 2.92 for GLM 5.3.23 The numbers are reported identically everywhere, but the framing differs a lot. Pandaily presented the result as strengthening WorkBuddy's position as a leading productivity agent in China.2 A more skeptical analysis noted that the margin is less than a tenth of a point on a 4-point scale, and that Tencent designed, ran and graded the test itself.11 I side with the skeptics. The result shows Hy4 preview belongs among the top Chinese open models. It does not show the model leads that group.

The public benchmark claims are bigger. Tencent reports 85.4 on Terminal Bench 2.1, ahead of DeepSeek V4 Pro. It also reports a jump on DeepSWE from Hy3's 28.0 to 64.3.4 One outlet says the Terminal Bench score matches Claude Opus 5.4 Other coverage stresses that every benchmark figure is still vendor-reported until someone reproduces it independently.18

The reporting also disagrees on some details. Most outlets describe the context window as exceeding one million tokens, while the official spec table lists 1M.12 Accounts of the predecessor also differ. One analysis places Hy3 preview in July 2026, while the family's Wikipedia entry dates Hy3 preview to April and the full Hy3 release to July 6.1114 These are small points, but they show how much of this story depends on Tencent's own materials.

The self-improvement claim

The most ambitious claim in the release is that Hy4 preview helped build itself. Tencent says the model took part for the first time in automatically optimizing training methods, data strategies, evaluation systems and low-level operators. According to the company, it proposed approaches, ran experiments and fed the results into later rounds, which Tencent calls an early loop of recursive self-improvement.2 Tencent also says the model found bottlenecks in its own inference system and improved operator fusion and communication, raising end-to-end throughput by 31.8% over baseline.8

The throughput gain is the easier part to test and the more useful one for anyone self-hosting. Because the weights and serving recipes are public, outside teams can check whether Tencent's inference efficiency holds up on their own hardware. The broader "recursive self-improvement" description is unverified, and readers should treat it as such.

Known limits and what comes next

Tencent is open about the model being unfinished. The repository says Hy4 preview tends to spend too long reasoning on complex tasks and to over-verify its own work, and that there is still room to improve both pre-training and post-training.1 Other coverage adds that it has no vision support.4 Tencent explicitly asks users to report what breaks.5

Pricing is $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached tokens.2 WorkBuddy and CodeBuddy were free for two weeks after launch.3 Since rebuilding its infrastructure in February, Hunyuan has released a major version about every two months, and Tencent says more Hy4 models are coming soon.2

The important change is that a frontier-scale Chinese model now comes with a permissive license, working deployment tooling and an outside safety score that is mostly encouraging. For security and infrastructure teams, the open weights mean Tencent's claims can be tested directly instead of taken on trust, and they should be tested before the model goes anywhere near production.