Tencent Hy4 Preview: The Biggest Apache 2.0 Model Yet, and What It Actually Scores

Tencent released Hy4 preview on August 28, 2026 and open-sourced the weights the same day. The headline number is 770 billion total parameters with 49 billion active per token, on a 1 million token context window. The number that actually matters is the licence: Apache 2.0. On the published record, Tencent Hy4 preview is the largest permissively licensed model anyone has shipped.

That distinction is doing more work than the parameter count. Kimi K3 is bigger at 2.8 trillion, but it carries a custom licence. GLM 5.3 published its own weights on exactly the same day under a bespoke Z.ai document rather than a standard one. Hy4 preview arrived under the plain, boring, lawyer-friendly Apache 2.0. Below are the verified specs, the benchmark scores with every source labelled, what the API costs, what hardware the official recipe assumes, and the weaknesses Tencent published about its own model.

The Key Takeaways

• Released and opened the same day: August 28, 2026, with weights on Hugging Face under Apache 2.0.
• Size: 770B total, 49B active per token, 78 layers, 256 routed experts plus one shared, 1,048,576 token context.
• Price: $0.834 per million input tokens and $2.501 output, with cached input at $0.042.
• Every benchmark score is Tencent’s own. There is no Artificial Analysis rating, and the single Arena placement is an automated one, not a human vote.
• You cannot run it: the official recipe assumes 16 B200s or 8 B300s, which is a data centre, not a workstation.

What Tencent Hy4 Preview Actually Is

Hy4 preview is the flagship of Tencent’s in-house model line, and the successor to Hy3. It is a mixture-of-experts language model aimed squarely at productivity work rather than chat: Tencent’s launch announcement names coding, office tasks, data analysis, game development and scientific research as the target areas, and describes generating a playable prototype from a single prompt inside Unity as one of the things it was built for.

One naming note, because it affects what you search for. Tencent has moved this line from “Hunyuan” to “Hy”, and both names remain in live use. The company writes “Tencent Hy Team”, the Hugging Face repository is tencent/Hy4-preview, and the GitHub organisation is still Tencent-Hunyuan. If you learned this family as Hunyuan, Hy4 is the same lineage under a shorter badge.

The word “preview” is Tencent’s own, and the model card treats it seriously rather than as marketing softening. This is described as an early version of Hy4, which is a meaningfully different claim from a finished flagship release.

Inside the Architecture

The specification is published in full on the model card, which is more than several recent Chinese releases have offered.

How the mixture-of-experts layout works

The backbone runs 78 layers. The first uses a standard dense feed-forward network, and the remaining 77 replace it with a mixture-of-experts block. Each of those blocks holds 256 routed experts and one shared expert. For every token the router picks the top 8 routed experts and adds the shared one, which is how a 770-billion-parameter model gets away with activating only 49 billion parameters at a time.

The practical consequence of that design is the one people consistently get wrong. A 49-billion active count does not make this a 49-billion parameter model to host. Routing selects a different subset for every single token, so all 770 billion weights have to be resident and reachable. The low activation count buys serving throughput once the model is loaded. It does not buy you a smaller download.

The 770B versus 780B discrepancy

Tencent says 770B everywhere. Hugging Face’s sidebar reads 780B params. The gap is not an error in either place, and the official vLLM deployment recipe resolves it: alongside the 770B total and 49B active, the model carries a 10B MTP layer for multi-token prediction. Hugging Face counts the checkpoint mechanically and includes it. Tencent quotes the backbone and does not.

This is the same shape of discrepancy that showed up when Z.ai shipped the GLM 5.3 weights at a counted 753B against a described 744B base. A mechanical parameter count and a vendor’s architectural description answer slightly different questions, and neither is evidence of a swapped model.

Tencent Hy4 Preview Benchmarks, With the Labels On

Here is where most coverage of this model goes wrong. Every score below comes from Tencent. No independent organisation has rated Hy4 preview, and the table says so in its own column rather than burying it in a footnote.

You will find tables elsewhere that place rival scores next to these, showing GLM 5.3 and DeepSeek V4 Pro trailing on SWE-Bench. Treat those with suspicion. The model card carries its comparison only as two images, benchmark.jpg and benchmark-appendix.jpg, with no machine-readable rival figures in the text. Numbers circulating for competitors on this release do not currently trace back to a primary source, so they are not repeated here.

What Tencent’s blind evaluation does and does not show

The comparison Tencent does publish in text is an internal one. The company ran 203 engineering tasks past 163 of its own experts in a blind rating exercise, scored out of 4.00. Hy4 preview averaged 2.99, against 2.94 for Kimi K3 and 2.92 for GLM 5.3.

Read that carefully. The spread between first and third place is 0.07 points on a four-point scale, judged by employees of the company whose model came first, on a task set that same company assembled. That is a reasonable internal signal and a poor public benchmark. It is not evidence that Hy4 preview beats Moonshot’s Kimi K3, and Tencent does not claim more than it is.

The one independent placement, and its asterisk

Arena posted a placement for Hy4 preview on its Code Arena WebDev board: roughly #5 overall at 1633 points, and about #3 among open models. That is a real jump, since Hy3 sat at #31 overall and #7 among open models.

The asterisk is large enough that Arena flagged it in the same announcement. This is an early AutoEval score, meaning a reward model trained on Arena’s human preference data cast the votes automatically instead of live human raters. It is a projection of how humans would probably vote, not a record of how they did. Until real votes accumulate, treat 1633 as provisional.

Pricing and How to Reach It

Tencent’s published rates are identical to the ones listed on OpenRouter, which is a useful consistency check on a launch-week price.

For flagship-tier capability that is aggressive pricing, and it undercuts most of what it is competing with at the frontier. It does not undercut the cheap end of the open field, where GLM 5.3 Flash sits at $0.15 and $0.50 under plain MIT. The two models are aimed at different jobs, and the gap in list price reflects that.

Where you can use it today

The API runs through Tencent Cloud TokenHub and through OpenRouter, where Tencent Cloud is currently the only provider serving the model. OpenRouter’s measured figures at launch put throughput around 35 tokens per second with a P50 latency of 3.90 seconds, and the context is listed at the full 1,048,576 tokens with completions capped at 64,000.

Inside Tencent’s own products, Hy4 preview is live in WorkBuddy and CodeBuddy in both domestic and international versions, and in Yuanbao and ima. Tencent offered the model free on WorkBuddy and CodeBuddy for two weeks from launch, which puts the end of that window in the second week of September 2026.

Can You Actually Run Tencent Hy4 Preview?

The weights are on Hugging Face under Apache 2.0, in a standard build and an FP8 variant. Downloading them is permitted, straightforward and, for almost everyone reading this, pointless.

What the official recipe assumes

The published model card and the vLLM recipe are unambiguous about the target hardware. The recommended configuration is 16 Nvidia B200s or 8 B300s, running the FP8 variant with --tensor-parallel-size 8. Serving is documented for vLLM and SGLang, both exposing an OpenAI-compatible endpoint with parsers for reasoning traces and tool calls.

That is a data centre deployment. Eight B300s is not a machine anyone has under a desk, and the mixture-of-experts routing described earlier means quantisation cannot rescue the situation the way it does for a dense 70B model. Self-hosting Hy4 preview makes sense when data placement, control or sustained heavy utilisation justify the operational cost. For everything else the API exists, and at $0.834 per million input tokens it will take a great deal of traffic before the hardware pays for itself.

Running it on a Mac

No. Not on an M5 Max, not on 256 GB of unified memory, not quantised. If your interest is open weights you can actually host on Apple silicon, the size class you want is one or two orders of magnitude below this, and our guide to the best open-source AI models covers what fits where. Reaching Hy4 preview from a Mac means calling it over the API like any other hosted model.

Why Apache 2.0 Is the Real Story

Three open releases closed August 2026 within three days of each other, and the licences separated them more than the scores did.

GLM 5.3 published its weights on August 28 under a bespoke document called the GLM 5.3 License, which tracks MIT closely but adds a clause requiring any model-as-a-service operator above $10 billion in revenue to pass a Z.ai security review first. Alibaba opened Qwen3.8-Flash-Next under the Qwen Community License 1.0 rather than Apache. Kimi K3, still the largest open model at 2.8 trillion parameters, ships under a custom Kimi K3 License.

Hy4 preview shipped under Apache 2.0, unmodified. No revenue threshold, no bespoke review clause, no company-specific terms for a lawyer to work through. For a legal team evaluating whether a model is usable in a commercial product, that difference outweighs a two-point benchmark gap, and it is the reason this release matters more than its provisional Arena rank suggests. Compared against how the rest of the field has been trending, including the licence and pricing shifts we tracked around DeepSeek V4, permissive terms at this scale are the genuine outlier.

What Tencent Says Is Wrong With It

The model card includes a short list of known issues, which is unusual enough to be worth quoting rather than paraphrasing. Tencent describes this as an early version of Hy4 with known problems, naming two: it spends longer than necessary reasoning through complex tasks, and it has a tendency to over-verify its own work.

Both are agentic failure modes rather than knowledge failures, and both cost money on an output-metered API. A model that re-checks work nobody asked it to re-check burns output tokens at $2.501 per million doing it. If you are budgeting a long-horizon agentic workload against Hy4 preview, price in more output tokens per completed task than the headline rate implies, and test the reasoning effort settings before committing to volume.

Is Tencent Hy4 Preview Worth Your Attention?

If you build products on open weights and answer to a legal team, yes, and the licence is the reason. Apache 2.0 at 770 billion parameters removes an obstacle that has been slowing adoption of every frontier-class Chinese model this year. That is a real change in the field, independent of where the scores land.

If you are choosing a model on measured quality today, wait. There is no Artificial Analysis rating, the only Arena number was cast by a reward model rather than by people, and every benchmark figure in circulation traces back to Tencent. None of that makes the scores wrong. It makes them unaudited, and a fortnight of independent evaluation will tell you more than the launch table does.

And if you wanted open weights you could run yourself, this is not that release. Eight B300s is the entry ticket. The download button is real, the licence is the most permissive in its size class, and the hardware behind it stays firmly out of reach.

FAQ