Tencent connects Hy models, AI workspaces and cloud model services

Tencent’s AI offer spans models, ready-made workspaces and a cloud platform for calling models. Tencent Hy is the current name of the family previously called Hunyuan; WorkBuddy puts agent capabilities into a workplace application; TokenHub provides a route for developers integrating models into their own systems. The useful buying decision is which layer the team needs. Choosing a model does not automatically choose an application, a hosting region or a commercial agreement.

01 / ProductThree routes into Tencent’s AI portfolio

Tencent’s international Hy3 announcement describes access through WorkBuddy, the Miora creative application, TokenHub and third-party developer platforms. This makes the portfolio relevant to both application users and engineering teams. It also creates a risk of comparing unlike offers: a workplace interface includes its own tools and experience, while a model API supplies a component from which a team must build those behaviours.

The Hy3 repository supplies a more concrete engineering reference. It describes a mixture-of-experts model, downloadable model variants and serving recipes using vLLM or SGLang. Its quick start demonstrates an OpenAI-compatible endpoint. That compatibility concerns the request interface; it does not establish identical reasoning, tool behaviour or output quality to another provider using a similar interface.

The release’s licence file states Apache 2.0. Read the licence attached to the exact downloaded release rather than applying that label to everything carrying the Hy name. Our analysis uses Hy3 as a documented deployment example. The current TokenHub price table also lists Hy4 preview, which should be treated as a separate preview candidate rather than a silent replacement in an established workflow.

02 / AudienceMatch the product layer to the team’s responsibility

An engineering team comparing language models can evaluate Hy through a hosted endpoint before deciding whether operating model weights is justified. A team already using Tencent Cloud may also value a common purchasing and operations environment. Neither circumstance establishes that it should move every workload: the relevant test is how this model handles the team’s own languages, documents and tool calls.

WorkBuddy is a different starting point for people who want an application to help with research, documents and routine office tasks. Application evaluation should examine what files and connected tools it can access, how actions are reviewed and what a user can export. A successful model experiment cannot substitute for those checks, because application permissions and model capability are separate properties.

For a model comparison, our DeepSeek blueprint provides another developer-facing reference. For a larger cloud portfolio, the Google blueprint helps frame the distinction between a model service and surrounding cloud products. Choose the comparison around the task and deployment constraints; company size alone does not determine which route is appropriate.

03 / WorkflowA proposed bilingual operations-assistant evaluation

Consider an operations team that receives English and Chinese supplier queries. It wants an assistant to locate an approved procedure, draft a response and identify exceptions that need a person. The following is a proposed evaluation, not a Tencent customer result or a test we performed. Its purpose is to separate language handling, evidence use and tool execution so failures have a clear owner.

Start with a small set of completed queries and their approved answers. Include ambiguous product names, conflicting delivery instructions and questions whose answer is absent from the available documents. Keep document versions with each example. An answer that sounds fluent but uses the wrong procedure should fail; a clear request for missing information can be the correct result.

For the China-site hosted route, Tencent’s TokenHub quick start requires a registered account and real-name verification, then service activation and an API key. It describes selecting the region and restricting the key’s accessible models or inference services. Confirm the suitable international account route separately if the team is outside that service context; the existence of a Chinese-language guide is not proof of universal eligibility.

Connect a deliberately narrow retrieval function to the trial application. Return the applicable procedure and its identifier, then ask the model for a structured draft containing the answer, evidence identifiers and an escalation reason if needed. Validate that each cited identifier came from retrieval. Do not let the model invent a new record identifier or decide that a missing policy can be replaced with a plausible instruction.

Run the same cases with the selected reasoning settings and record the model version, full request cost, completion time and human corrections. Review both languages with someone who can judge the operational meaning. Back-translation can expose obvious discrepancies, but it should not be the sole acceptance test: two translations can preserve the same original mistake.

If the assistant later gains a tool that changes a delivery instruction, separate drafting from execution. Have the user confirm the proposed change and bind that confirmation to the exact supplier record and values. Store the tool result independently of the assistant’s summary. A timeout must lead to checking whether the change already occurred before a retry can create a duplicate action.

Only consider self-hosting after the hosted experiment establishes useful behaviour and there is a reason to own serving operations. The repository’s deployment recipes are a starting reference for infrastructure planning, not evidence of performance on the team’s hardware. Test model loading, concurrent demand and failure recovery with the actual serving stack before committing a production workflow.

04 / PricingKeep the region, model and billable unit together

The TokenHub price table was readable in a public browser on 16 September 2026. The examples below come from its China-site Guangzhou language-model table and use CNY per million tokens. They are not universal dollar prices. A quote or console selection for another region, provider or application must be checked against that specific route.

Tencent’s billing guide separates input, output and cached-input charges. It also says generated tokens can be billed when an inference request is interrupted. A user cancelling a long answer therefore should not be modelled as a free request. Different media models use other units and rules; do not apply a language-token estimate to video, image or speech workloads.

For a practical cost model, count the complete query: retrieval context, previous conversation, generated answer and any repeated attempt. Then divide total spend by the number of answers accepted after review. Keep that operating measure separate from the published token tariff. It reveals whether a cheaper request merely moves effort into retries or staff correction.

05 / DistinctionsTencent offers several ways to place the same model work

The distinguishing feature is the connection between a model family, applications and cloud delivery. A team can ask whether its problem is best solved by adopting an existing workspace, integrating an API or operating a release itself. Those choices carry different responsibilities for permissions, updates and support, even when they draw on related model research.

Tencent publishes model architecture and serving information alongside product announcements, allowing a technically capable buyer to investigate more than a marketing demo. Its benchmark and productivity claims remain company-reported. We have not reproduced them, and they should not be copied into a business case as expected savings. The supplier-query experiment provides a more relevant local test.

06 / QuestionsResolve version and access before comparing outcomes

The biggest open question is the exact offer available to the intended account. Tencent’s international announcement, China-site documentation and third-party hosting routes describe overlapping but different access paths. Confirm model identity, data-processing region, retention and tool support in the selected service. A familiar model label is not enough to establish that all providers expose the same configuration.

The other question is whether the assistant can preserve an instruction across languages and several turns. Test a user correcting a delivery location after the first draft and a later question referring back to that correction. Measure whether the proposed action stays attached to the right record. This is more decision-relevant than a broad claim of conversational fluency.

07 / DecisionChoose one Tencent route and prove the task

Tencent deserves consideration as an established model developer and cloud provider with application options. Begin with the smallest route that answers the actual need, preserve the exact version and price basis, and expand only after the workflow produces acceptable results. Keep a fallback for unavailable models and an explicit handoff for missing evidence.