Tencent Hy4 Preview is available on Novita AI as a serverless text-to-text API. The exact model ID is tencent/hy4-preview. As of September 20, 2026, Novita AI lists $0.834 per million input tokens, $2.501 per million output tokens, $0.042 per million cached input reads, a 1,000,000-token context window, and a 64,000-token maximum output. The model is a 770B-parameter Mixture-of-Experts system with 49B active parameters, and Novita AI positions it for coding agents, complex tool-use workflows, planning, context continuity, and sustained multi-step execution.
| Tencent Hy4 Preview on Novita AI | Details |
|---|---|
| Model name | Hy4 Preview |
| Model ID | tencent/hy4-preview |
| Context window | 1,000,000 tokens |
| Max output | 64,000 tokens |
| Input price | $0.834 per 1M tokens |
| Output price | $2.501 per 1M tokens |
What Is Tencent Hy4 Preview?
Hy4 Preview is Tencent’s preview-class Hy-generation model for coding agents, complex tool use, and productivity work. The Novita AI model page identifies it as a Mixture-of-Experts model with 770B total parameters and 49B active parameters. That split matters for API users because it points to a large-capability model where fewer parameters are active on each token, rather than a dense model where every parameter contributes to every request.
The preview label carries real engineering meaning. A preview model can still change before a full release, so it should be evaluated as a strong candidate for experimentation and incremental rollout rather than adopted immediately as a stable production default. If you build against it, version your prompts, record model behavior alongside task outcomes, and re-run evaluation when the model changes.
This is a text-only chat model. It is not an image, audio, or video model, and it does not replace specialized retrieval or embedding models.
Tencent Hy4 Preview API Access on Novita AI
Novita AI lists Hy4 Preview as a serverless model with chat/completions, anthropic, and responses endpoint support. The current feature list includes function calling, structured outputs, reasoning, and serverless delivery.
The most direct integration is the OpenAI-compatible chat/completions path. Point your client at Novita AI’s API base URL, set the model field to tencent/hy4-preview, and authenticate with a Novita AI API key. If you already run an OpenAI-compatible agent or application, this is mostly a model-routing change rather than a protocol migration.
The Novita AI model page also includes a playground, which is useful for validating model routing before you put a new request into a larger evaluation.
| Access item | Novita AI value |
|---|---|
| Display name | Hy4 Preview |
| Exact model ID | tencent/hy4-preview |
| Model type | Chat |
| Access mode | Serverless |
| Endpoint families | chat/completions, anthropic, responses |
| Input modality | Text |
| Output modality | Text |
| Features | Serverless, function calling, structured outputs, reasoning |
Hy4 Preview Specs and Pricing on Novita AI
The model specification and pricing below use the Novita AI model listing checked on September 20, 2026.
| Spec / pricing item | Novita AI value |
|---|---|
| Model ID | tencent/hy4-preview |
| Status | Live, serverless |
| Architecture | Mixture-of-Experts |
| Total parameters | 770B |
| Active parameters | 49B |
| Context window | 1,000,000 tokens |
| Maximum output tokens | 64,000 tokens |
| Input price | $0.834 per 1M tokens |
| Output price | $2.501 per 1M tokens |
| Cached input read price | $0.042 per 1M tokens |
| Created in catalog | September 17, 2026 |
Cached input reads are priced at $0.042 per 1M tokens when your request reuses qualifying input context. This rate applies to cached input reads rather than to every input token.
Capabilities: What Hy4 Preview Can Do
Hy4 Preview is designed for multi-step work. Novita AI’s description explicitly targets coding agents, complex tool-use workflows, and productivity tasks that need planning, context continuity, and sustained execution. In an API context, that makes the listed capabilities more relevant than headline size:
- Function calling lets the model choose a controlled application tool and emit arguments for that tool.
- Structured outputs help move model decisions into downstream systems without relying entirely on fragile natural-language parsing.
- Reasoning support gives you another axis to evaluate for difficult planning, code analysis, or workflow execution tasks.
- A 1M input window gives room for broad context, though you still need retrieval discipline and explicit token budgets.
Long context does not remove the need for context engineering. Put task instructions, policy constraints, current repository state, retrieved documents, and relevant tool results into a deliberate order. Keep old failures, irrelevant logs, and unrelated code out of the request. For a coding agent, the useful test is not whether the model can accept 1,000,000 tokens, but whether the request contains the right context to make a correct change.
Developer Fit: When to Use Hy4 Preview
Use it for coding-agent and tool-use evaluations. If your current model loses track of multi-file context, emits invalid tool arguments, or cannot sustain a plan through several tool turns, Hy4 Preview is a useful comparison point. Test it on real failing tickets, dependency upgrades, refactors with tests, and workflows that require multiple tool calls rather than single-turn snippets.
Use it for long-context analysis. The 1,000,000-token window fits requests that combine specification, repository context, architecture notes, issue history, logs, and evaluation constraints. This is most valuable when keeping all relevant state in one request produces a better decision than chunking alone.
Use it when structured handoffs matter. Structured outputs and function calling make it practical to connect the model to a review queue, issue tracker, test runner, or deployment workflow, provided your application validates tool arguments and result formats.
Use it when cacheable context repeats. Stable system instructions, tool schemas, repository summaries, and policy blocks may benefit from the cached input-read rate. Evaluate actual cache behavior and cost across representative calls instead of assuming the discount applies uniformly.
When Not to Use It
Short, low-cost tasks. A small classifier, routing job, short summarizer, or low-stakes rewrite usually does not need a 770B/49B active MoE model. Keep a cheaper model as the baseline unless quality has already proven to be the bottleneck.
Vision, audio, or video workloads. Hy4 Preview is text-in and text-out on Novita AI. Use a dedicated multimodal model if screenshots, UI video, audio, or image understanding is central to the workflow.
Hard latency floors. MoE routing and deep reasoning workloads can behave differently from dense chat models under load. Measure time to first token, end-to-end latency, retries, and tool-cycle time before routing interactive traffic to a new model.
Repetitive, stable workflows with no evidence of quality loss. If your existing pipeline passes tests and keeps failure rates low, switch only when a representative evaluation shows a real gain, not because a newer model is available.
How to Evaluate Hy4 Preview Against Your Current API
Start with a fixed evaluation set and keep the scaffold identical across models. The same tools, retrieval setup, timeout, output cap, and grading rules should apply.
For coding agents, collect task-level outcomes, not only response impressions:
- Did the change pass the relevant tests?
- Did the agent call the correct tools in a valid order?
- How many retries, corrections, or human interventions were needed?
- What were the total input, cached input, and output tokens?
- Did latency remain acceptable for the product workflow?
For automation, track schema validity, downstream acceptance, correction rate, and cost per completed task. Those numbers make it easy to decide whether Hy4 Preview should become the default, remain an escalation path, or stay out of production.
A practical rollout policy is to keep Hy4 Preview as a specialist for difficult coding-agent steps or long-context planning while your existing model continues to handle high-volume, low-risk traffic. Promote it only when its measured advantage outweighs the higher output price and the operational cost of changing a preview dependency.
Conclusion
Tencent Hy4 Preview on Novita AI is worth evaluating for coding agents, tool-heavy automation, and long-context planning. The launch facts are concrete: tencent/hy4-preview, serverless access, 770B/49B MoE, a 1,000,000-token context window, a 64,000-token output ceiling, and $0.834/$2.501 input/output per million tokens as of September 20, 2026.
Its fit is not universal. The preview label, text-only inputs, and higher output cost mean it should prove itself on your workload. Start with the Hy4 Preview API and Playground, run the same tasks you use for your current model, and compare correctness, tool reliability, latency, and total token cost before promoting it.
FAQ
What is the model ID for Tencent Hy4 Preview on Novita AI?
The exact model ID is tencent/hy4-preview.
Is Tencent Hy4 Preview available on Novita AI?
Yes. Novita AI lists Hy4 Preview as a live serverless chat model as of September 20, 2026.
How much does Tencent Hy4 Preview cost on Novita AI?
As of September 20, 2026, it costs $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input reads.
What is the context window and maximum output?
The Novita AI listing shows a 1,000,000-token context window and a 64,000-token maximum output.
Does Hy4 Preview support function calling and structured outputs?
Yes. The Novita AI listing includes function calling, structured outputs, reasoning, and serverless support.
What are the active and total parameter counts?
Novita AI describes Hy4 Preview as a 770B-parameter Mixture-of-Experts model with 49B active parameters.
Is Hy4 Preview the same as Hunyuan?
The article uses Tencent’s current Hy generation display name. It is not related to Hunyuan Image or Hunyuan Video, which are separate media-generation product families.
