Hy4 vs Hy4 Preview: What’s Documented for Coding, Tools, and Structured Output
Compare the available evidence for Hy4 and Hy4 Preview, and use a privacy-first checklist before handling user data or credentials.
Summary · TL;DR
Tencent describes Hy4 Preview’s coding capabilities, but the supplied source does not establish a capability comparison with Hy4 or document a tool-calling interface or structured-output guarantees. Teams handling sensitive data should verify the exact deployment and test with synthetic inputs before considering production use.
Key Takeaways
- Tencent documents software-engineering capabilities for Hy4 Preview, including long-context development, debugging, and validation.
- The supplied evidence does not establish whether Hy4 and Hy4 Preview are the same model, different versions, or how their capabilities compare.
- Tool-calling interfaces, schema-constrained output, and privacy guarantees remain unverified in the supplied source.
- Keep credentials out of prompts and tool results; validate outputs and authorize tool actions outside the model.
On this page (7)
The decision: evaluate Hy4 Preview, but don’t assume feature parity
If your workflow handles user data or credentials, treat Hy4 and Hy4 Preview as an unverified comparison. Tencent’s Hy4 Preview release describes coding and productivity capabilities for Hy4 Preview. It does not establish how a separately listed Hy4 model compares, or document a tool-calling interface, schema-constrained output, or data-handling guarantees.
OneMux, the publisher of this article, lists Hy4 and Hy4 Preview separately. Those listings are useful for finding model access, not evidence that the two entries are identical or different in capability. Verify the exact model and deployment before sending any sensitive material.
Hy4 vs Hy4 Preview: capability evidence
The table separates Tencent’s statements about Preview from questions the supplied evidence cannot answer. A description of a capability is not an independent measurement or a guarantee for every API deployment.
| Question | Hy4 Preview | Hy4 comparison |
|---|---|---|
| Coding | Tencent describes understanding, planning, debugging, and validation for long-context development, plus front-end development improvements. | Unknown. The supplied source does not compare a separately listed Hy4 model with Preview. |
| Tool calling | Tencent describes the model running experiments and working with game engines, but does not specify an API tool-calling interface or its reliability. | Unknown. Neither parity nor a performance difference is established. |
| Structured output | Tencent describes creating documents, spreadsheets, and presentations; it does not document JSON-schema enforcement or valid-output rates. | Unknown. No supported comparison is available. |
| Long context | Tencent states a context window exceeding 1 million tokens for Preview. | Unknown from the release as a comparison; do not infer effective retrieval or privacy properties from window size. |
| Privacy and credentials | The supplied release does not specify retention, training use, regional processing, or credential-handling guarantees. | Unknown; check the terms for the exact provider and deployment. |
Source: Tencent’s Hy4 Preview release. OneMux lists both models with a 1M context value, but that catalogue detail does not establish equivalent behavior or a version relationship.
What Tencent’s coding evaluation does—and does not—show
Tencent reports an internally conducted blind evaluation involving 163 experts and 203 engineering tasks. It reports an average score of 2.99 out of 4.00 for Hy4 Preview, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. These are Tencent-reported results for that evaluation, not proof of general superiority or a measurement of Hy4 versus Hy4 Preview. The supplied summary does not provide enough task-level detail to establish performance on credential-sensitive code. See the Tencent release.
For a practical decision, test the coding work you actually need: for example, whether the model can identify a secret accidentally logged by an application, propose a fix, and preserve existing behavior. Use synthetic credentials and review the proposed patch manually. A plausible explanation is not evidence that the patch is secure.
A privacy-aware evaluation workflow
The following is an illustrative, untested checklist, not a claim about either model’s security. Start with a non-production environment and fabricated data.
- Identify the exact service path. Record the model identifier, endpoint, service provider, and any intermediary receiving requests. OneMux offers an OpenAI-compatible API for model access; compatibility with a particular tool or structured-output parameter remains conditional on documented endpoint and parameter support.
- Check data terms before use. Obtain current, deployment-specific answers about request and response retention, training use, access controls, subprocessors, and processing location. The Tencent release supplied here does not answer these questions.
- Minimize the input. Replace names, account identifiers, tokens, and passwords with consistent synthetic placeholders. Send only the code or records needed for the task. A large context window is not a reason to upload an entire repository or user database.
- Keep secrets outside the model. Store credentials in an appropriate secret-management system. Do not include them in prompts, conversation history, tool descriptions, tool results, or error messages returned to the model.
- Test the boundaries. Use fabricated records to check whether outputs repeat sensitive-looking strings, whether logs capture prompts or responses, and whether a proposed action exceeds the user’s authorization. Inspect both application logs and any available provider controls.
- Approve production use separately. Have the appropriate security and privacy reviewers assess the actual deployment, data classification, contracts, and operational controls. A successful synthetic test does not establish compliance or eliminate leakage risk.
For example, instead of sending a real access token when asking for a logging fix, substitute EXAMPLE_TOKEN_NOT_A_SECRET and show only the relevant logging function. Check whether the suggested patch removes the value from logs without breaking required diagnostics.
Tool calls and structured output: gates before automation
Tencent’s account of Preview running experiments is useful evidence of a workflow, but it is not an API specification. Before designing an agent around either catalogue entry, confirm that the exact endpoint and model document the tool-call parameters, response format, and any restrictions you intend to use. If support is undocumented, treat it as unknown rather than assuming it exists.
When tool calling is supported, keep authorization outside the model: validate tool arguments, grant the tool only the permissions it needs, and require human approval for sensitive or irreversible actions. Treat tool output as untrusted input; it can contain private data or instructions that should not override your application’s rules. Avoid returning raw credentials to the model even when a tool needs credentials internally.
The same distinction applies to structured output. A model may produce JSON when asked without guaranteeing that it matches your schema. If the chosen endpoint documents schema-constrained output, verify its limitations; otherwise, parse and validate responses in your application. Reject invalid or unexpected fields, and do not execute an action merely because the response is syntactically valid. Use bounded retries with sanitized error messages, then send failures for review rather than silently accepting malformed data.
Conclusion
Hy4 Preview has Tencent-documented coding capabilities, but the available evidence does not support a Hy4-versus-Preview winner for coding, tool calling, or structured output. For privacy-aware work, the decisive next step is to verify the exact endpoint and data terms, then test with synthetic inputs and external safeguards before considering real user data.
Sources
- Tencent: Tencent Releases and Open-Sources Tencent Hy4 Preview — model description, reported evaluation, and productivity claims.
FAQ
Is Hy4 the same model as Hy4 Preview?
The supplied evidence does not establish whether they are identical, different versions, or otherwise related. OneMux lists them separately, which by itself does not prove a capability difference.
Does Hy4 Preview support tool calling?
Tencent describes workflows in which Hy4 Preview ran experiments and worked with game engines, but the supplied release does not specify an API tool-calling interface. Check documentation for the exact model and endpoint.
Does Hy4 Preview guarantee valid JSON or schema-compliant output?
The supplied Tencent release does not document such a guarantee. Check endpoint-specific documentation and validate outputs in your application.
Can I safely send credentials to Hy4 or Hy4 Preview?
The supplied source does not provide a credential-handling guarantee. Keep credentials out of prompts and tool results, and review the data terms and controls for the specific deployment before using sensitive data.
Related articles
Guides
Hy4 vs. Hy4 Preview for API Incident Diagnosis: What Is Documented?
Compare the available evidence for Hy4 and Hy4 Preview, and use a practical checklist to evaluate coding, tool calling, and structured output for API incident diagnosis.
Guides
Hy4 Flash vs Hy4 Preview for production workflows: what is documented?
Compare Tencent's documented claims about Hy4 Flash and Hy4 Preview, then design a workflow with validation, bounded retries, and safe recovery.
Guides
DeepSeek V4 Flash vs GPT and Qwen: Which Model Should You Use?
A practical comparison of DeepSeek V4 Flash, GPT, and Qwen for coding, production workflows, cost control, and model selection through OneMux.
Guides
Grok 4.6 vs Claude Opus 5: Coding Cost, API Pricing, and the Practical Pick
Compare Grok 4.6 and Claude Opus 5 for coding workloads, token economics, and deployment through OneMux’s OpenAI-compatible API.