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.
Summary · TL;DR
Tencent reports that Hy4 Flash slightly outperformed Hy4 Preview overall on its internal WorkIndex assessment at a lower per-task cost. The supplied source does not establish a head-to-head comparison for coding, tool calling, or structured output, so production teams should validate those requirements and build recovery rules around observed failures.
Key Takeaways
- Tencent recommends Hy4 Flash for tasks including document processing and simple front-end displays, but the source does not provide a Flash-versus-Preview coding benchmark.
- Native tool calling and schema-constrained output are not established for either model by the supplied source; unknown does not mean unsupported.
- Use bounded retries for transient failures, validate every actionable output, and require reconciliation before retrying a possibly completed side effect.
- Tencent's WorkIndex cost and quality comparison is an internal, aggregate result, not a guarantee for an individual workflow.
On this page (10)
The decision: start with the workflow, not a capability assumption
Hy4 Flash is a reasonable candidate for everyday production workflows involving documents, presentations, or simple front-end displays. Tencent reports that it was slightly better overall than the larger Hy4 Preview on its internal WorkIndex assessment, with one-thirteenth of Preview's cost per task. That is an aggregate vendor result—not proof that Flash is better at coding, tool calling, or structured output. Tencent's announcement does not supply those head-to-head measurements.
For an operational workflow, first establish whether the endpoint supports the features you need. Then test the entire path from model response to validation, action, and recovery. This article is published by OneMux, which provides access to both models through a unified API; the operational guidance also applies if you use another provider.
Hy4 Flash vs Hy4 Preview: documented facts and gaps
The following comparison is limited to Tencent's announcement. A gap means the supplied source does not establish the claim, not that the model lacks the capability.
| Question | Hy4 Flash | Hy4 Preview | What you can conclude |
|---|---|---|---|
| Coding | Tencent recommends Flash for simple front-end displays. | The source does not give a comparable coding result. | A front-end recommendation is not a coding benchmark or evidence that Flash is stronger. |
| Tool calling | Native tool-calling interface and performance are not specified in the supplied source. | Not specified in the supplied source. | Support and comparative reliability are unknown. |
| Structured output | Schema-constrained output support and performance are not specified in the supplied source. | Not specified in the supplied source. | Do not assume either model enforces a JSON schema. |
| Everyday work assessment | Tencent reports a slight overall advantage on WorkIndex. | Used as the larger reference model in Tencent's comparison. | This is an internal aggregate result, not a result for every task type. |
| Cost per task | Tencent reports one-thirteenth of Preview's cost on its WorkIndex calculation. | Reference point for that calculation. | Actual cost depends on the workload and provider prices. |
WorkIndex covers tasks such as presentation creation and modification, document editing, spreadsheet repair, audiovisual processing, and visual understanding. Tencent also recommends the separate Hy4 model—not Hy4 Preview—for complex work such as long-horizon software engineering. That recommendation should not be converted into a claim about Preview's coding ability.
Tencent lists Flash at 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active; it lists Preview at 770 billion total and 49 billion active. Parameter counts alone do not establish which model will handle a particular workflow better.
Prerequisites for an operational trial
Before connecting either model to a production action:
- Confirm the endpoint contract. Check the current provider documentation for model availability, authentication, request limits, and any native tool-calling or structured-output parameters. An OpenAI-compatible API does not by itself establish support for every OpenAI-style parameter on every model.
- Define the output contract. Specify required fields, allowed values, size limits, and what makes an output safe to act on. Validate these in your application even if an endpoint offers constrained output.
- Separate reading from writing. Identify which operations only retrieve data and which create, modify, send, or delete something. Give action credentials only the permissions the workflow needs.
- Prepare recovery state. Record a workflow identifier, current step, validated input, action identifier, and outcome. Avoid storing sensitive payloads unnecessarily.
- Create a review path. Decide who receives a failed or ambiguous job and what information they need to resolve it.
OneMux provides access to Hy4 Flash and Hy4 Preview through its OpenAI-compatible API. Verify endpoint and parameter support before relying on a particular integration; no integration is claimed as tested here.
An illustrative workflow with retry and recovery rules
Consider a workflow that reads a support operations document, extracts proposed changes, and prepares a ticket update. This is a design example, not a tested result or a claim that either model supports a native tool interface.
1. Generate a proposal, not an immediate action
Ask the model for the ticket identifier, proposed status, and a short justification. Treat the response as untrusted text. Parse it and validate the identifier format, required fields, and allowed status values in application code. Check the ticket against the authoritative system rather than trusting an identifier invented in the response.
If parsing or validation fails, retain the failure reason and make at most one corrective request that includes the validation error. Do not repeatedly resubmit an unchanged prompt. If the corrected response still fails, route the job to review. If schema-constrained output is available, use it as an additional control, not a substitute for business validation.
2. Retry transient failures with a limit
For a timeout or temporary service error before an action starts, retry with a bounded attempt count and increasing delay; for example, allow up to three attempts, subject to your latency budget. Respect any retry guidance returned by the provider. Authentication errors, invalid requests, and permission failures should stop the workflow for correction rather than trigger automatic retries.
Record each attempt and its outcome. Alert or queue the job when the retry budget is exhausted instead of silently discarding it.
3. Reconcile uncertain side effects
A timeout after submitting a ticket update does not prove the update failed. Before retrying, query the ticket system for the action identifier or expected state. If the system supports idempotency keys, use a stable key for the same intended action. If the outcome cannot be determined safely, pause for human review; a blind retry could duplicate a comment or repeat another side effect.
For destructive or externally visible actions, add an approval gate appropriate to the risk. A valid model response is not authorization to act.
How to compare the models on your own workflow
Use the same representative, permission-safe cases and acceptance criteria for both models. Include ordinary cases, missing information, malformed source documents, ambiguous instructions, and attempted prompt injection. Measure separately:
- Coding, if relevant: whether generated changes pass your tests and review, not merely whether code is produced.
- Tool use, if supported by the endpoint: whether the model selects an allowed operation and supplies valid arguments. Keep execution behind application authorization.
- Structured output: parse success, schema validity, and business-rule validity as distinct checks.
- Operational results: completed tasks, review rate, retry rate, end-to-end time, and total cost, including failed attempts.
Use a held-out set for the final comparison and inspect failures before expanding access. The result may differ from Tencent's WorkIndex: your tools, documents, acceptance rules, and cost structure are different.
Conclusion
Tencent's source supports an aggregate everyday-work comparison in Flash's favor, but it does not settle Flash versus Preview for coding, native tool calling, or structured output. Start with a bounded trial, verify the endpoint contract, and make validation and recovery part of the workflow—not an afterthought.
Sources
FAQ
Is Hy4 Flash better than Hy4 Preview for coding?
The supplied Tencent source does not provide a head-to-head coding benchmark. It recommends Flash for simple front-end displays, but that does not establish comparative coding performance.
Does Hy4 Flash support tool calling and structured output?
The supplied source does not establish native tool-calling or schema-constrained output support. Check the documentation for the specific endpoint and model before designing around those features.
Should a workflow retry an action after a timeout?
Not blindly. If an action may already have completed, first reconcile its outcome with the target system. Use a stable idempotency key where supported, or send an unresolved outcome to human review.
Related articles
Guides
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.
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
Qwen4 Max for Production Workflows: What Can You Compare With Qwen3.8 Max?
Separate documented Qwen4 Max capabilities from unknowns, then design a workflow evaluation with explicit retry and recovery rules.
Guides
Hy4 Flash API Costs: A Reproducible Workload Estimate
Estimate Hy4 Flash API costs with explicit billing-unit assumptions, cache scenarios, and a repeatable workload evaluation plan.