Tencent releases Hy4 preview with open weights and a 1M-token context

Tencent's Hy4 preview is a 770B-parameter MoE model with 49B active parameters, a 1M-token context, Apache 2.0 weights, and official serving recipes.

Quick answer

Tencent released Hy4 preview on August 28, 2026 with model weights and supporting code under Apache 2.0. Tencent describes a 770B-parameter Mixture-of-Experts backbone with 49B parameters active per token, a 1M-token context window, and a separate native MTP layer for speculative decoding. Official vLLM and SGLang recipes expose an OpenAI-compatible local endpoint. This is an early preview with acknowledged over-reasoning and over-verification issues, and its benchmark and productivity comparisons are first-party results. Availability in Chat AI has not been verified.

Download Chat AI Opens the official App Store or Google Play for your device.

Quick answer: Hy4 preview is a very large open-weight MoE model

Tencent has released Hy4 preview, an instruction-tuned Mixture-of-Experts model aimed at coding, office analysis, game development, and scientific research. Its official architecture table lists a 770B-parameter backbone, 49B parameters active per token, and a 1M-token context window. Tencent published the model weights, an FP8 variant, serving recipes, and a fine-tuning pipeline. The release is explicitly a preview rather than the finished Hy4 family.

Sources: Tencent, Tencent Hy on GitHub, Tencent on Hugging Face

The 770B figure covers the backbone, with 49B active for each token

Hy4 preview has 78 backbone layers. Tencent says 77 of them use 256 routed experts plus one shared expert; each token activates eight routed experts and the shared expert. The model card lists 770B total backbone parameters and 49B activated parameters. It also describes a separate native multi-token-prediction layer with 10B total parameters and 0.7B active, designed for speculative decoding. That distinction is useful when storage tools or hosting pages report a packaged parameter count above the 770B backbone figure.

Sources: Tencent Hy on GitHub, Tencent on Hugging Face

Tencent published base and FP8 weights under Apache 2.0

The official repository links an instruction model and an FP8 version on Hugging Face, ModelScope, GitCode, and CNB. The repository and model card identify Apache 2.0 as the license. Tencent also says the model can be used through WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter. Those product routes, prices, and launch offers can change independently of the downloadable weights, so users should confirm current terms at the point of access.

Sources: Tencent, Tencent Hy on GitHub, Tencent on Hugging Face

Self-hosting is documented, but this is not a typical consumer-PC model

Tencent's quickstart uses vLLM or SGLang to serve Hy4 preview behind an OpenAI-compatible API. The official examples use the FP8 weights with eight-way tensor parallelism and specialized attention, reasoning, tool-call, and speculative-decoding settings. Tencent does not present that recipe as a small-device configuration. Teams evaluating local deployment should calculate weight storage, accelerator memory, key-value cache, context length, and concurrency for their own hardware before assuming that open weights mean easy desktop use.

Sources: Tencent Hy on GitHub

The productivity comparison is an internal Tencent evaluation

Tencent reports a blind side-by-side evaluation in which 163 internal experts rated outputs on 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. The repository also includes a larger benchmark appendix. These are useful first-party release measurements, but they are not an independent ranking and do not establish a permanent winner across coding, documents, science, price, latency, or safety. A fair adoption test should use the team's own prompts and review criteria.

Sources: Tencent, Tencent Hy on GitHub

Tencent says Hy4 helped optimize parts of its own development stack

According to Tencent, Hy4 preview participated in experiments involving training methods, data strategies, evaluations, and low-level operators. The company also reports that model-led work on operator fusion and communication optimization increased end-to-end inference throughput by 31.8% against its baseline. This is a provider-reported result from Tencent's development environment, not proof that the model can autonomously improve itself without human-defined systems, measurements, code review, or infrastructure access.

Sources: Tencent

The preview has known over-reasoning and over-verification issues

Tencent's model card says Hy4 preview can spend longer than necessary reasoning through complex tasks and may over-verify its work. The default reasoning mode is high, while the serving template also provides a no-think option for direct responses. Those controls do not remove the need to test instruction following, latency, tool use, factuality, and output length on representative tasks. Preview status also means behavior, weights, recipes, and access routes may change as Tencent develops later Hy4 models.

Sources: Tencent Hy on GitHub, Tencent on Hugging Face

Hy4 preview availability in Chat AI is not verified

Hy4 preview is not present in Chat AI's current verified model catalog, so this article does not claim that it can be selected in the app. Chat AI users can still apply the same evaluation discipline to models that are available: compare the same task, inspect factual support, record latency and failure modes, and avoid treating a provider benchmark as a substitute for real workflow testing.

Sources: Chat AI, Tencent, Tencent Hy on GitHub

Frequently asked questions

What readers usually ask

Is Tencent Hy4 preview open source?

Tencent published the Hy4 preview weights, repository, deployment recipes, and a fine-tuning pipeline under the Apache 2.0 license. Users should still review the license and each linked dependency before commercial deployment.

How large is Hy4 preview?

Tencent lists a 770B-parameter Mixture-of-Experts backbone with 49B parameters active per token. The model card separately lists a native MTP layer with 10B total parameters and 0.7B active for speculative decoding.

What context length does Hy4 preview support?

The official model card lists a 1M-token context window. Effective context in a real deployment still depends on serving configuration, available memory, concurrency, and task quality.

Can Hy4 preview run locally?

It can be self-hosted with the official vLLM or SGLang recipes, but Tencent's FP8 examples use eight-way tensor parallelism. The release does not present Hy4 preview as a normal consumer-laptop model.

Does Hy4 preview beat GLM-5.3 and Kimi K3?

Tencent reports a small lead in its internal blind evaluation of 203 engineering tasks. That first-party result does not prove Hy4 preview is better for every workload, deployment budget, latency target, or safety requirement.

Is Hy4 preview available in Chat AI?

Availability has not been verified. Hy4 preview is absent from Chat AI's current verified model catalog, so users should not assume it can be selected in the app.

Evidence

Sources

  1. Tencent Hy4 preview release announcementTencent · Primary source
  2. Hy4 preview official repository and deployment guideTencent Hy on GitHub · Primary source
  3. Hy4 preview model card and weightsTencent on Hugging Face · Primary source
  4. Chat AI verified model directoryChat AI · Primary source