Tencent Hy4 and Hy4 Flash: What Changes After Preview?
Tencent says Hy4 and Hy4 Flash are coming soon. Compare vision, long-horizon tasks and reasoning efficiency with Preview, plus API availability, prices and three practical tests.
Published 2026-10-29
Tencent announced on October 29, 2026 that Hy4 and Hy4 Flash are coming soon. At its Global Digital Ecosystem Summit, Dowson Tong described stronger long-horizon productivity, improved reasoning efficiency and visual understanding for full Hy4. Flash inherits vision and agent capabilities with a focus on value. The conference report gives no exact launch date or API price.
Checked 2026-10-29: APIMaster has 0 active public routes for Hy4, Flash and Preview. Qwen below is a separate available alternative. USD per 1M input / output tokens:
| Model ID / filter | Official price¹ | APIMaster current price | Active routes | Discount | Model card |
|---|---|---|---|---|---|
hy4 |
Not announced | No public route | 0 | — | Check |
hy4-flash |
Not announced | No public route | 0 | — | Check |
hy4-preview |
Hosted quote unverified | No public route | 0 | — | Check |
qwen3.8-flash |
$0.15 / $0.47 | From $0.06 / $0.18 | 6 | 60% off input / ~62% off output | Available |
¹ Qwen uses Alibaba Cloud's International tier. Its Global tier is $0.113 / $0.382; savings against that tier are about 47% / 53%. The market's configured reference is $0.15 / $0.45; we use the checked vendor price for the comparison. Caching and tools are excluded. The marketplace card is the source of truth for APIMaster charges. hy4 and hy4-flash are filters, not confirmed API IDs.
GPT Trial(GPT-6 Astra, GPT-6 Sol, GPT-5.6, GPT-5.5, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-6.1-sol, gpt-6-astra, gpt-6-sol, and gpt-6-luna available)
Sign up and claim a $20 GPT trial
Hy4 vs Preview: three meaningful changes
Vision is the clearest new capability. Preview's official card describes text generation. Full Hy4 adds visual understanding, and Flash inherits it. Screenshots, charts and visual evidence can become part of a workflow. This is the Hy4 generation's first vision upgrade, not Tencent's first vision model; it does not announce image generation or audio/video APIs.
Long tasks should produce usable deliverables. Preview already targets coding, office documents, games and research. The announcement emphasizes stronger long-horizon and real productivity tasks. A useful test is whether a model preserves requirements through multiple steps, recovers from a tool error and finishes a working artifact with fewer human corrections.
Reasoning efficiency addresses an acknowledged weakness. Tencent explicitly lists unnecessarily long reasoning and over-verification among Preview's limitations. The new efficiency claim directly addresses that problem. Measure repeated checks, output tokens and time at the same acceptance standard. No percentage gain or serving-speed figure has been published in the report.
Preview was released on August 28, 2026 with 770B backbone parameters, 49B active per token, a 1M context and Apache 2.0 weights. Those are Preview specifications; full Hy4 and Flash need their own parameter, context and licensing disclosures. The article does not present unreported benchmarks.
Hy4 Flash: inherited vision and agents, value-oriented positioning
Tencent says Flash will evolve through close model/product Co-Design, inheriting Hy4's vision and agent abilities. It has not supplied Flash pricing, parameter count or latency.
A practical evaluation strategy is to try Flash on routine extraction and first-pass visual checks, then test full Hy4 on ambiguous evidence and longer dependency chains. This is a suggested workload split, not a measured ranking. Compare cost per accepted task, including retries and tools: a cheaper token rate alone does not establish a cheaper completed job.
Three practical tests for release day
These are proposed prompts, not official demos or completed Hy4 tests. Preserve the same inputs and scoring. Compare Preview only on shared text tasks; visual comparisons require image-capable routes.
Screenshot to interface fix: “Given this screenshot and these component files, identify three layout defects, propose the smallest patch and verify the result at a 390px viewport. Distinguish visible evidence from checks that require running the page.” Score fixed defects, regressions and manual corrections.
Chart and spreadsheet reconciliation: “Reconcile this chart image, CSV and amendment. Return the corrected revenue table with chart-label and row citations. Flag missing evidence.” Score numbers, units and resolution of the deliberate contradiction.
Interrupted long task: “Implement these requirements in the supplied repository, recover from the reproducible tool failure, run checks and deliver the patch plus verification report.” Track completion, repeated tool calls, output tokens, elapsed time and total cost across multiple runs.
Hy4 API pricing and access
Full/Flash API IDs, prices and rollout details were not specified in the announcement. Watch the Hy4 card and Flash card. Preview's self-hosted hy4-preview example does not establish hosted full-model access.
For an immediate baseline, Qwen3.8-Flash has 6 active routes, ranging from $0.06–$0.15 input / $0.18–$0.45 output per 1M tokens, checked 2026-10-29. Minimum rates save 60% / about 62% against International pricing, or about 47% / 53% against Global. This is a different model, not a Hy4-equivalence claim.
Register → create an API key → use https://apimaster.ai/v1 with qwen3.8-flash for the available baseline. Keep model selection configurable and verify image/tool support on the selected route. Once Hy4 and Flash open publicly, APIMaster will integrate them promptly and offer the most competitive pricing available.
FAQ
Are Hy4 and Hy4 Flash released?
Tencent announced on October 29, 2026 that they are coming soon. That is not confirmation of public API access. Preview is already released.
What changes from Hy4 preview?
Visual understanding, stronger long-horizon productivity and more efficient reasoning were announced. Numerical gains have not been provided in the checked report.
What are the API IDs and prices?
Full/Flash identifiers and prices remain unspecified in the announcement. APIMaster currently has no public route; the lowercase names here are availability filters.
Are the new models open source with a 1M context?
Apache 2.0 and 1M are documented for Preview. The upcoming variants need their own release documentation.