Hy4-preview API for Coding and Long-Horizon Agents

Hy4 preview is a text-to-text flagship model released by the Tencent Hy Team on August 28, 2026. It is also searched as Tencent Hy4, Tencent Hunyuan 4, Hunyuan4 API, and Hy4-preview API, but the official public checkpoint is named Hy4 preview. GPTProto provides hosted API access to the model, while Tencent separately publishes downloadable weights for teams that want to operate their own inference stack.

The model uses a Mixture-of-Experts architecture with 770 billion backbone parameters and 49 billion activated for each token. Its 78-layer backbone contains one dense feed-forward layer followed by 77 MoE layers. Each MoE layer has 256 routed experts and one shared expert, with eight routed experts selected per token. A separate native MTP layer supports speculative decoding.

Tencent's model card lists a 1M-token context length and describes Gated DeepSeek Sparse Attention with IndexCache for sparse index reuse. This preview checkpoint accepts text and returns text; it does not natively accept images, audio, or video.

Hy4-preview API Applications

Repository-scale coding: Give a coding agent enough context to inspect architecture, interfaces, tests, logs, and related files before proposing a patch. Tencent specifically highlights improvements in planning, debugging, verification, and front-end implementation quality.

Long-horizon agent workflows: Use the model for tasks that require repeated planning, execution, inspection, and revision. The surrounding agent application should still control tools, permissions, timeouts, retries, and acceptance checks rather than treating model output as automatically verified.

Document and data work: Analyze information distributed across long documents, produce structured summaries, and support workflows that create documents, spreadsheets, presentations, equations, or financial analyses.

Game development and research: Prototype game logic through multi-turn iteration or assist with difficult research problems in areas such as AI development, molecular dynamics, condensed-matter physics, and mathematics. Outputs should be reviewed against domain evidence before use.

Hy4-preview Upgrades and Known Preview Trade-offs

Tencent describes Hy4 preview as a major generation change driven by increases in model size, context length, training data, and post-training. The company also co-designed the model with products such as CodeBuddy and WorkBuddy, giving the release a stronger focus on end-to-end productivity rather than short, isolated benchmark prompts.

The word preview matters. Tencent reports that this early checkpoint can reason longer than necessary on difficult tasks and may over-verify its own work. For short classification, extraction, or formatting requests, compare latency and token use against a smaller model. For long coding or agent runs, test task completion, tool-call validity, generated-token volume, and recovery from failed steps before sending production traffic.

Hy4-preview vs GLM-5.3 and Kimi K3

Tencent ran a blind internal evaluation in which 163 experts reviewed outputs for 203 engineering tasks. Hy4 preview received a 2.99/4.00 average score, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3.

These figures were reported by Tencent and were not independently reproduced by GPTProto. They suggest that Hy4 is competitive on Tencent's engineering task set, not that it wins every coding workload. A fair API comparison should reuse the same repository snapshot, system prompt, tool definitions, token budget, and pass/fail checks across models.

Hosted Hy4-preview API vs Self-Hosting the Open Weights

Open weights do not make inference operationally free. Tencent's reference deployment uses specialized vLLM or SGLang configurations for sparse attention, MTP speculative decoding, reasoning parsing, and tool-call parsing. Hosted access removes that serving work, while self-hosting remains the better option when infrastructure control is more important than setup time.

When Should You Choose Hy4-preview?

Choose Hy4 preview when your application benefits from large-context code understanding, cross-document analysis, sustained planning, or access to an Apache 2.0 open-weight flagship without running the checkpoint yourself. It is especially relevant for developers evaluating Chinese frontier models for coding and agent workloads through a shared API account.

Use a smaller or more mature model when requests are short, latency-sensitive, or highly repetitive. Because this is a preview release with known over-reasoning and over-verification behavior, validate cost, response time, task success, and output consistency on your own workload before making it the default model.