Hy4 Preview FAQ Specifications, API, Pricing, Regions, and Limits

v_cloudbuilder
2026-08-281596

This FAQ answers the product questions international developers are likely to ask before evaluating Hy4 Preview. Pricing, availability, regions, and Preview conditions are time-sensitive. The facts below were checked on August 28, 2026; verify current Tencent Cloud International documentation before integration.

View Hy4 Preview on Tencent Cloud International TokenHub and check its current access details

What is Hy4 Preview, and when was it released?

Hy4 Preview is an early version in Tencent Hunyuan’s Hy4 iteration. Tencent’s release page is dated August 28, 2026. It positions the model for software engineering, office analysis, game development, and scientific research, while the international campaign summarizes agentic, coding, and productivity scenarios. Positioning is not a guarantee for every workload.

Is Preview the same as a stable production release?

No. Tencent describes Hy4 Preview as an early version and says pre-training and post-training still have room to improve. Preview should not be presented as GA or unconditionally production-ready. Organizations need a scoped PoC, acceptance thresholds, human escalation, and a rollback before limited live traffic.

What do 770B and 49B mean?

Tencent reports 770B total parameters and 49B activated parameters. They describe different aspects of the architecture. The activated count is not the total model size, and neither number alone proves application accuracy, latency, infrastructure requirements, or economic value.

How should 1M context, 960k input, and 64k output be read?

Tencent Cloud International lists a 1M context window, 960k maximum input, and 64k maximum output. Calling 1M the maximum input is inaccurate. Practical requests need room for instructions, tool results, and output, and remain subject to current API and account constraints. Capacity does not guarantee reliable retrieval from every position in a long prompt.

What model ID should an API request use?

The listed model ID is hy4-preview. Query GET /v1/models against the intended regional endpoint and verify the returned availability status before invoking it. Preview availability can change, so a stale sample or screenshot is not sufficient operational evidence.

Which capabilities are listed for Hy4 Preview?

The international model list marks Deep Reasoning (Preserved Thinking), Structured Output, Function Calling, and Caching. These labels need application controls: validate structured fields, authorize and execute tools server-side, and observe actual cache accounting. Preserved thinking does not establish factual correctness, and the listing does not document every possible protocol option.

What is the Singapore price?

As checked on August 28, 2026, the Singapore pricing page lists input at US$0.834 per 1M tokens, output at US$2.501 per 1M tokens, and cache hit at US$0.042 per 1M tokens. These are regional, time-bound US-dollar facts. Do not combine them with mainland-China RMB pricing or assume they apply in Silicon Valley forever.

How should a team estimate task cost?

Use the actual billed quantities for ordinary input, cache hits, and output. Include every turn, retry, returned tool result, and regeneration in the business task. A cache hit must not be assumed. Compare estimates with console usage, and track user operations separately from attempts so retries do not disappear from the cost model.

Can you show a hypothetical Singapore calculation?

Yes, if every quantity is clearly labeled as an example assumption. Suppose actual metering shows 1M ordinary input tokens, 0.5M cache-hit tokens, and 0.1M output tokens. Using the prices checked on August 28 gives US$0.834 + US$0.021 + US$0.2501, or US$1.1051. This is not a quote or a cache-hit promise. Longer context, retries, and tool-result messages change the eventual bill.

What is the Singapore endpoint?

The default Singapore domain is https://tokenhub-intl.tencentcloudmaas.com; the versioned Base URL is https://tokenhub-intl.tencentcloudmaas.com/v1. Documentation labels its resource scheduling scope Global. “Global” does not mean unconditional sale in every country, absence of network restrictions, or automatic compliance with local data rules.

What is the Silicon Valley endpoint?

The default Silicon Valley domain is https://tokenhub-us.tencentcloudmaas.com, with Base URL https://tokenhub-us.tencentcloudmaas.com/v1. Its resource scheduling scope is United States. That label does not by itself prove state-level residency, a certification, or low latency from every US location. Confirm account activation, credentials, quota, and billing for that region.

Are backup API domains available?

The API instructions list .tech backup addresses for Singapore and Silicon Valley to try when a default address is unavailable. A backup hostname is not a normal traffic-balancing target or proof of automatic cross-region recovery. Authentication failures, exhausted quota, and an offline model are not fixed by changing DNS names. Do not blindly replay a request with an unknown execution state, especially when tools may have produced external side effects.

Are OpenAI and Anthropic API formats supported?

Tencent Cloud International documents compatibility directions for OpenAI and Anthropic API protocols. Compatibility is not behavioral identity. Model coverage, parameters, tool semantics, thinking fields, streaming events, errors, and SDK behavior may differ. Test the exact model, path, region, and client version required by the application.

How does API authentication work?

OpenAI-compatible access uses Authorization: Bearer <API_KEY>. Inject keys from a secret manager and never expose them in browsers, repositories, or logs. Use the regional GET /v1/models call to check authentication and model status. A 401 should trigger credential diagnosis, not uncontrolled retries.

Is Hy4 Preview open source?

Tencent says Hy4 Preview is open source and links to GitHub, Hugging Face, ModelScope, and GitCode/AtomGit entry points. Read the current repository license, scope of released weights and code, version notes, and known issues before use. Open source does not automatically mean zero cost, unrestricted commercial use, or no operational burden.

What did Tencent’s internal blind test report?

Tencent says 163 internal experts scored 203 engineering tasks on a four-point scale, with Hy4 Preview averaging 2.99/4. This is Tencent’s internal blind test, bounded by its tasks, model versions, raters, and procedure. It is not an independent public benchmark and cannot establish comprehensive superiority. Enterprise buyers should repeat blind evaluation on their own tasks.

What pairwise figures did Tencent publish?

Tencent reports GLM 5.3 at 2.92/4 and Hy4’s win/draw/loss against it at 46.8%/12.8%/40.4%. It reports Kimi K3 at 2.94/4, with Hy4 at 51.2%/7.9%/40.9% in that pairwise comparison. An average score and pairwise win/draw/loss answer different questions. Neither should be converted into a universal ranking beyond the internal test’s versions, tasks, and process.

What known behavior has Tencent disclosed?

Tencent notes that complex tasks may involve longer-than-needed thinking and excessive self-verification, and that this early model has room for improvement. Applications should cap steps, tool calls, elapsed time, tokens, and cost, then stop loops that add no evidence. The behavior is not stated to occur on every request, and no universal incidence rate is published.

Does the long context mean every file should be uploaded at once?

No. The context specification is a capacity boundary, not a content-selection policy. Duplicate records, stale versions, and irrelevant attachments can crowd out useful evidence or create contradictions. Inventory sources, bind versions, select material relevant to the question, require resolvable citations, and test retrieval at different lengths and evidence positions. A larger prompt is not automatically a better prompt.

What should a production rollback cover?

A rollback is more than selecting another model name. Preserve the last accepted model and prompt versions, protocol mapping, output validator, and tool-permission policy. Decide how in-flight streams and tool operations are reconciled so switching cannot duplicate an external action. Trigger rollback on predefined quality, error, latency, cost, or safety thresholds, and test it during the PoC rather than for the first time during an incident.

Can an enterprise use Hy4 Preview in production?

Preview availability alone is not production approval. Evaluate representative and adversarial tasks for quality, citations, structure, latency, cost, permissions, failure recovery, and data obligations. Require human approval for high-impact actions and retain versioned telemetry. If internal gates pass, begin with a small canary and a tested route back to an already accepted model. Recheck pricing, regions, status, and terms before each release.

Create. Connect. Grow.

Share your ideas, showcase your story, and get support from our team anytime.