Shift log: Hunyuan 4 on Chinese support drafts

The support queue needs a draft that an agent can approve without repairing the policy, the tone, or the promised outcome. For Chinese customer support, fluent wording is only the first checkpoint. The harder requirement is preserving the distinction between a customer request, a verified account fact, and an action the company has authorized.

This shift note puts Hunyuan 4 head-to-head with Qwen3.8 Max for that drafting job. It does not claim measured results: no shared test set or model outputs were supplied. The comparison therefore establishes decision rules and deployment verdicts rather than inventing a winner.

Hunyuan 4 and Qwen3.8 Max get the same ticket packet

Give both models identical customer messages, order facts, policy excerpts, and response instructions. Require a customer-facing draft in Simplified Chinese and a separate internal note identifying missing information. Keep this article and the evaluation rubric in English; assess the actual drafts with reviewers who handle Chinese-language support.

Use resolved, de-identified tickets covering refund requests, delayed shipments, subscription cancellations, damaged goods, and complaints that require escalation. Include incomplete records and conflicting customer statements. Those cases reveal whether a model fills gaps with plausible but unsupported details.

Run the first pass without model-specific prompt tuning. Then allow the same tuning effort for each candidate. Otherwise, the exercise measures which prompt received more attention, not which system better fits the desk.

Refund disputes make policy fidelity the first gate

Start the head-to-head with a refund request whose eligibility remains unresolved. Both models should acknowledge the inconvenience, explain the next verification step, and avoid promising approval. A polished answer that invents an exception loses to a plainer answer that stays inside policy.

• Reject either draft if it invents a refund amount, processing deadline, eligibility decision, or completed action.
• Check whether each draft separates confirmed facts from conditions that still require an agent or system check.
• Prefer the draft that gives the customer a concrete next step without shifting unnecessary work onto them.

The verdict for refund drafting is policy fidelity before style. Neither Hunyuan 4 nor Qwen3.8 Max should advance merely because its language sounds more reassuring.

Angry customers expose tone and escalation mistakes

Next, use a complaint containing repeated contacts and a demand for compensation. Compare whether each model acknowledges the specific failure without repeating a generic apology. Reviewers should assess natural Chinese phrasing, appropriate formality, and whether the response sounds like an accountable support agent rather than promotional copy.

Keep escalation rules explicit. If policy requires specialist review, the draft should explain the handoff without implying that compensation is already approved. If no escalation route is provided, the model should flag that gap internally instead of inventing a department or callback arrangement.

The verdict for complaint handling is conditional. Favor whichever candidate consistently combines appropriate tone with accurate escalation boundaries. Treat tone as a reviewer judgment, not an assumed advantage attached to either model name.

Agent editing time decides the operational winner

Blind the model identities during review and record whether drafts are accepted, lightly edited, heavily rewritten, or rejected. Track unsupported commitments separately from stylistic edits. Measure generation latency and actual billed usage in your environment; neither a low token price nor a fast response alone establishes lower cost per approved draft.

For a mixed-model desk, Relay365 is a practical option to evaluate as a pay-as-you-go relay. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Its API base is https://api.api365relay.com; OpenAI clients add /v1. Confirm current model availability, pricing, and data-handling terms before submitting support records.

Create a Relay365 account if that shared integration fits your evaluation workflow, then begin with de-identified tickets and a capped test budget.

Shift verdict: pilot Hunyuan 4, retain the challenger

• Hunyuan 4 earns a supervised pilot, not an evidence-free production endorsement.
• Qwen3.8 Max remains the head-to-head challenger until blinded review establishes a meaningful operational difference.
• Choose the model with fewer policy violations and less agent rework, using cost and latency to distinguish otherwise acceptable candidates.
• Keep human approval for refunds, compensation, account changes, and unresolved policy questions regardless of which model wins.