Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

Tencent launches lower-cost Hy4 flash for everyday AI work

Tencent’s November 16 Hy4 flash release adds a lower-cost hosted model with image understanding, tool calling and a million-token context.

Published 16 Nov 2026· 4 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

Tencent launched Hy4 flash on November 16, 2026, adding a lower-cost AI model to its hosted lineup. TechNode reports the launch, and Tencent Cloud’s updated documentation lists the model, its API prices and a million-token context. For university teams, the release creates another candidate for evaluating routine document and software tasks.

A new Flash tier joins Tencent’s model lineup

The release follows Tencent’s November 4 introduction of Hy4. That earlier announcement positioned the full model for coding, professional work and scientific research, while identifying a Flash version as a forthcoming, more cost-efficient option.

The new model now has its own hy4-flash identifier in Tencent Cloud’s API guide, updated November 16. The guide lists support for OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols. It separately lists Hy4 and marks Hy4 preview as approaching retirement.

TechFlow’s November 16 report, citing Jinshi Data, also reports the Flash launch and its million-token context. ArtificialWatch’s same-day account identifies it as the cheaper tier and links the release to Tencent’s announcement and TokenHub pricing. These reports distinguish the new Flash release from the earlier flagship introduction.

What the official specification supports

Tencent Cloud’s model list gives Hy4 flash a one-million-token context window, a maximum input of 960,000 tokens and maximum output of 64,000 tokens. It lists image understanding, structured output, function calling, caching and deep reasoning with preserved thinking.

The same specification gives the full Hy4 model the same listed context, input and output limits. That makes context capacity a poor basis for choosing between the two by itself. Matching capacity does not establish matching answer quality, task completion or reliability; those require evidence from the task being performed.

For research groups, CASRAI’s assessment is that the combination is relevant to document-heavy workflows and tools that need structured responses. A model evaluation suite should still test whether the chosen model follows the required format and grounds its answers in the supplied evidence. A documented feature is a starting point for evaluation, rather than proof that a complete workflow works correctly.

Lower token prices are the concrete purchasing change

Tencent Cloud’s published pricing table lists Hy4 flash at $0.07 per million input tokens, $0.281 per million output tokens and $0.003 per million cached-input tokens. Its Singapore table lists full Hy4 at $0.834, $2.501 and $0.042 respectively.

These are separate input, output and cache-hit categories. A project with long inputs and short answers has a different billing mix from one producing lengthy outputs. Comparing only the cheapest cache rate would therefore misdescribe the ordinary input cost of the model.

CASRAI’s editorial assessment is that research offices comparing hosted tools should record the applicable region, rate categories and actual token use for a representative task. A lower unit price can justify running an evaluation; whether it lowers the cost of an acceptable result depends on the result’s accuracy and the amount of review or repetition it requires.

Compatibility does not remove integration requirements

The API guide explains that TokenHub uses model-specific service activation and credentials. It says a model’s service must be enabled in the console before calls can succeed, and distinguishes Singapore and Guangzhou credentials. The documentation also warns that supported capabilities and individual parameters vary by model.

For a university software team, that means protocol compatibility should not be treated as an automatic, complete replacement for an existing integration. CASRAI’s assessment is to test tool calls, structured responses, error handling and account configuration together. The test should cover the application’s intended permissions, rather than only a successful chat response.

Recording the model identifier and relevant settings also supports computational reproducibility. This is especially useful where a preview and a released model remain listed at the same time: describing both simply as “Hy4” would obscure which service produced the output.

What the launch establishes, and what comes next

TechNode’s headline cost reduction refers to Tencent’s internal task benchmark. The official pricing table establishes token rates, while a benchmark measures a particular set of tasks. Those are different kinds of evidence. CASRAI’s assessment is that institutions should compare completed work and review requirements before adopting a vendor’s percentage saving as their own budget assumption.

The immediate next step for teams considering the service is a bounded evaluation of the released hosted model. Research offices can connect that evaluation to their AI governance practices, identifying the permitted data and the person responsible for reviewing outputs. The sources cited here establish a release and documented capabilities, rather than an institutional approval to process any particular dataset.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →