Hy4 preview has documented API access. Tencent’s August 28 release announcement names Tencent Cloud TokenHub and OpenRouter as access routes. The earlier version of this guide incorrectly treated a missing entry in our catalog as proof that no provider offered it.
Match the release first
Tencent identifies the preview as a mixture-of-experts model with 770 billion total parameters and 49 billion active parameters. Those are vendor specifications, not measurements by this site. They replace the earlier speculative comparison with unrelated models.
Keep Hy4 preview, Hy4 and Hy4 flash separate in your comparison sheet. A price, promotion or benchmark for one release does not establish the same terms for another. Record the provider’s exact model identifier, region, service tier and date alongside every quote. Check the model catalog for tracked offers, then verify the chosen offer with the provider. A missing row means our comparison is incomplete.
Calculate a bill you can check
For uncached text, use input and output rates from the same provider:
cost = input_tokens / 1,000,000 × input_rate
+ output_tokens / 1,000,000 × output_rate
For an illustrative workload of 10 million input and 2 million output tokens, this becomes 10 × input_rate + 2 × output_rate. This is a workload example, not a Hy4 price quote. Add cache, tools, minimum fees and other billable categories when they apply. Keep a dated copy of the provider terms so a later pricing change does not invalidate your comparison silently.
A temporary application trial is not necessarily a production API allowance. Verify the account’s quota, expiration, billing currency and rate limits before budgeting. Compare accepted task results as well as token bills; fewer retries may offset a higher token price, but only an evaluation on your workload can establish that.
Self-hosting requires separate evidence
Active parameter count alone cannot size resident weights. Obtain the actual checkpoint, license, supported runtime and precision before checking GPU offers. Memory for runtime buffers and concurrent contexts must also fit. This guide does not certify a GPU configuration or predict Hy4 throughput.
Use the cost crossover worksheet once you have both a real API quote and measured serving capacity. Do not infer an API price, quality level or safe deployment size from another model’s parameter count.
Reviewed November 18, 2026. This replacement removes the unsupported no-API claim, speculative architecture bracket and provider-spread predictions.