Tencent Hy Hy4 Preview

Overview

Tencent Hy4 Preview is the August 28, 2026 preview release of Tencent's next-generation flagship language model. It moves the Tencent Hy line from the Hy3 generation to a larger open Mixture-of-Experts model with 770 billion total parameters, 49 billion active parameters per token, and a 1M-token context window.

The release is not just a larger checkpoint. Tencent positions Hy4 Preview around real productivity tasks, including software engineering, office analysis, game development, finance, security, and scientific research. The main decision for Hy3 users is whether the larger context window, Apache 2.0 weights, stronger productivity training, and new hosted routes justify moving to an early preview that Tencent says still has known reasoning and over-verification issues.

What's New

Scales the Hy Flagship to 770B Parameters

Hy4 Preview expands the flagship model scale from Hy3's 295B total / 21B active MoE profile to 770B total / 49B active parameters. The model uses 78 backbone layers, with dense feed-forward computation in the first layer and MoE layers afterward. Each MoE layer includes 256 routed experts and one shared expert, with the top eight routed experts activated per token.

The practical implication is a larger open model aimed at harder long-horizon work. Tencent frames the release as a step change from Hy3, driven by model size, context length, and training data rather than a small product update.

Adds a 1M-Token Context Window

Hy3 and Hy3 Preview were already long-context models at 256K tokens. Hy4 Preview raises that ceiling to 1M tokens, which changes the evaluation for codebases, research corpora, financial models, multi-file office workflows, and agent traces.

The model architecture uses Gated DeepSeek Sparse Attention with IndexCache for sparse index reuse across layers. For buyers and builders, the key point is simpler: Hy4 Preview is the current Tencent Hy release to test when 256K context is not enough.

Uses Productivity-Focused Training and Evaluations

Tencent says Hy4 Preview was trained with data co-created by internal experts in software engineering, game development, finance, security, and related domains. It is designed for work that produces artifacts, not only chat answers: documents, spreadsheets, presentations, playable prototypes, code changes, and research analysis.

Tencent also reports a blind internal evaluation with 163 experts across 203 engineering tasks. In that evaluation, Hy4 Preview averaged 2.99 out of 4.00, slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. Treat those as vendor-run internal results, useful for positioning but not a substitute for your own workload tests.

Changes the License and Deployment Story

One of the clearest deltas from Hy3 is licensing. Hy4 Preview is released under Apache License 2.0, while the earlier Hy3 release used Tencent's custom Hy Community License. That makes Hy4 Preview materially easier to evaluate for teams that require a permissive open-source license before commercial adoption.

Tencent publishes both Hy4 Preview and Hy4 Preview-FP8 weights through Hugging Face, ModelScope, GitCode, and CNB. The GitHub README includes vLLM and SGLang deployment recipes, OpenAI-compatible API examples, fine-tuning guidance, quantization references, and recommended inference settings. Self-hosting is still a serious infrastructure project, especially for a 770B MoE model, but the deployment route is more explicit than a simple model-card launch.

Compared With Hy3

Hy4 Preview is the current exploratory flagship. Hy3 remains the safer reference when you need the official July 2026 release rather than a preview checkpoint.

Choose Hy4 Preview when you need the 1M context window, want Apache 2.0 model weights, or are evaluating the next Tencent open frontier model. Stay on Hy3 when preview stability, lower operational risk, or an already validated deployment matters more than the larger model.

Availability & Access

Hy4 Preview is available through Tencent products including WorkBuddy, CodeBuddy, Yuanbao, ima, and other Tencent surfaces. Tencent says users can also connect through Tencent Cloud TokenHub and OpenRouter.

At launch, Hy4 Preview is free on WorkBuddy and CodeBuddy for two weeks. Tencent also extended free Hy3 access on both platforms until September 30, 2026. Those free periods are time-limited launch terms, so production users should verify availability and quotas before making cost assumptions.

For self-hosted testing, Tencent publishes Hy4 Preview and Hy4 Preview-FP8 model weights on Hugging Face, ModelScope, GitCode, and CNB. The official README recommends vLLM or SGLang for production serving and provides prebuilt image examples for both runtimes.

Pricing & Plans

Tencent's launch announcement lists API pricing for Hy4 Preview at:

These prices are release-announcement figures, not a full international billing matrix. Confirm the exact provider, region, currency, quota, free period, and cache behavior in Tencent Cloud TokenHub or OpenRouter before estimating production cost.

Known Limitations

Tencent explicitly describes Hy4 Preview as an early version. The official README says there is remaining headroom in pre-training and post-training, and it names two known issues: the model can spend longer than necessary reasoning through complex tasks, and it may over-verify its own work.

That matters for agent workflows. Extra reasoning and repeated verification can improve confidence on hard tasks, but they can also increase latency, token usage, and tool-call volume. Teams should test Hy4 Preview on real prompts before replacing a stable Hy3 deployment.

Who Should Upgrade / Who Should Wait

Upgrade or test Hy4 Preview if:

• You need a 1M-token context window for codebases, research sets, large documents, or long agent traces.
• You require Apache 2.0 weights before approving commercial self-hosting or model customization.
• You are building productivity agents for coding, office artifacts, finance, game development, or scientific work.
• You want to compare Tencent's current open flagship against GLM-5.3, Kimi K3, or other large reasoning models.

Wait or stay on Hy3 if:

• You need a non-preview model with a more stable operational profile.
• Your workloads already fit comfortably inside Hy3's 256K context window.
• You cannot absorb extra latency or token usage from a model that may over-reason or over-verify.
• You need fully documented provider availability, quotas, and enterprise billing before migration.

FAQ

Sources

• Tencent Hy4 Preview launch announcement
• Tencent-Hunyuan Hy4 Preview GitHub repository
• Hy4 Preview README