Tencent Hy4 Review: Vision, Coding, Computer Use — November 2026

Tencent Hy4 is most interesting when it reduces the gap between understanding material and doing something useful with it. Whether it deserves your money depends heavily on the product and billing route you choose. An office user may value fewer transfers between documents and spreadsheets. A developer needs to know whether a completed, checked task still costs less after tool calls, retries, and review.

Our initial recommendation is a targeted trial for people already working in Tencent productivity tools, or teams considering a long-context model service. The buying decision needs more care than the launch headlines suggest. Image support, personal subscriptions, custom applications, and local deployment have different boundaries. A million-token window or a computer-use demonstration does not, by itself, justify replacing a workflow that already works.

1. What Hy4 is, and who makes it

Hy4 is a general-purpose language and visual-understanding model developed by the Hunyuan team at Tencent, a Chinese technology company. The Hunyuan family made its public debut on September 7, 2023. Hy4 preview followed on August 28, 2026; the official Hy4 release was announced on November 4, 2026. These are the family debut, the preview release, and the new model release, respectively.Tencent: Hunyuan public debut in 2023 · Tencent: Hy4 preview announcement · Tencent: Hy4 launch announcement

For purchasing purposes, separate the model from the place where you use it. Yuanbao, ima, WorkBuddy, and CodeBuddy are applications or working environments. TokenHub supplies model services. A model interprets inputs and generates outputs; an application supplies files, accounts, tools, permissions, and a way to inspect the result. Two products using the same model can therefore offer substantially different experiences.

That distinction matters more than it first appears. A knowledge-work user needs reliable access to existing material and clear references. A coding user needs changes that can be checked inside a real project. An office user needs an editable document rather than a convincing description of one. Missing permissions, an unsuitable execution environment, or a poor export format can erase much of the benefit of a stronger model.

Tencent emphasizes visual understanding, computer use, and collaboration between model development and product development. Our editorial reading is that the relevant question is how much of a task can remain in one coherent workflow. Does the tool keep the necessary context, perform the right action, and leave an inspectable result? Treating Hy4 as a universal membership covering every Tencent AI feature would obscure those practical questions.Tencent: Hy4 launch announcement

2. Versions, identifiers, and documented limits

Tencent Cloud lists the official model as hy4 and the preview as hy4-preview. They remain separate entries. We did not obtain an official statement that requests naming hy4-preview are automatically upgraded to Hy4. An existing integration should therefore establish which model it actually uses. A shortened label in an application is not enough evidence that its underlying model has changed.Tencent Cloud: models and documented limits

Tencent Cloud: models and documented limits · Tencent: Hy4-preview model card and license · Tencent: Hy4 launch announcement · Tencent Cloud: personal Token Plan subscriptions

The pay-as-you-go catalog lists a 1M-token context window, a maximum input of 960k tokens, and a maximum output of 64k. It also lists structured output, function calling, caching, and thinking capabilities. Read the individual limits alongside the overall context constraint. Adding the maximum input and maximum output figures together does not establish that both can be filled simultaneously in every request.Tencent Cloud: models and documented limits

A large context window can accommodate project instructions, selected code, meeting notes, and earlier decisions in the same task. Its practical value depends on how well the model uses that material. Conflicting document versions and poorly identified sources can still produce an unreliable answer. For long-document work, locating evidence and explaining missing information often matters more than filling every available token.

The preview was disclosed as a 770B-parameter model with 49B active parameters. We keep those figures attached to the preview rather than assigning them to the official release. Tencent says Hy4 weights and a system card will follow. At our check, we had not obtained a verifiable official-release weight package and full license. Hardware planning and commercial deployment terms need the documentation for the exact release.Tencent: Hy4 preview announcement · Tencent: Hy4-preview model card and license · Tencent: Hy4 launch announcement

3. What vision, coding, and computer use could change

Vision: bring screenshots and charts into the discussion

Image understanding is explicitly listed for Hy4 in the model-service catalog. A reasonable use case is discussing a product screenshot, a chart, or a document that mixes text and graphics. Instead of describing every visible detail manually, the user can provide the visual material. Useful analysis still requires the model to understand units, legends, time ranges, and any relevant context that is absent from the image.Tencent Cloud: models and documented limits

Consider a dashboard combining cumulative totals and monthly results. A useful response would distinguish those definitions, identify unreadable fields, and explain which visible evidence supports its conclusion. This is an illustrative acceptance standard, not a test we ran. Low-resolution images, small labels, and dense legends make it particularly valuable to keep the original source available for checking.

Image understanding also needs to be separated from image generation. TokenHub lists image and video generation as other model categories. A Hy4 endpoint that accepts an image should not automatically be described as producing finished posters or videos. An application can combine several models behind one interface, but capability and billing still depend on what it actually invokes.Tencent Cloud: models and documented limits

Coding: examine the finished change, not just its first draft

A coding task includes understanding the requirement, editing the project, running appropriate checks, and explaining the change. A polished first draft can still omit empty states, mishandle updates, or assume the wrong dependency version. For a model intended to work with tools, the useful outcome is a checked implementation that respects the project constraints. A screenshot of an attractive page measures only a small part of that outcome.

Tencent describes optimization for front-end development, games, and professional software tasks. Those directions are worth investigating, but we do not turn demonstrations into production pass rates. A static landing page and an administration interface with authentication, pagination, and several roles require different evidence. Existing naming conventions, build systems, test environments, and external services can all change how much work remains after generation.Tencent: Hy4 launch announcement

Computer use: capability requires an execution environment

The announcement describes computer use across command-line and graphical interfaces. An application or agent framework must still provide screenshots, actions, a terminal, access to applications, and the necessary authorization. Buying a model service does not automatically grant control of an entire computer. Which applications can be accessed, how a task can be stopped, and what the user can inspect are product-level questions.Tencent: Hy4 launch announcement

The most relevant tasks are sequences: read material, organize a table, prepare a draft, or modify a project and run its existing checks. Correctly clicking one button is only an intermediate step. Window changes, expired sessions, duplicate filenames, and tool errors can interrupt the sequence. Tasks with inspectable outputs and reversible actions are a useful starting point for deciding whether delegation actually saves effort.

4. Published evaluations: encouraging signals, limited conclusions

Early launch coverage can blur the difference between a vendor evaluation and an independent conclusion. We use two before-and-after metrics from Tencent’s internal search evaluation. We do not reproduce the launch announcement’s competitor ranking or combine unrelated benchmarks into a synthetic league table. Keeping a comparison within the same metric makes its meaning easier to assess.

Tencent reports a hallucination rate falling from 12.95% to 7.2%, and a guessing-error rate falling from 23.1% to 5.6%. Those changes are 5.75 and 17.5 percentage points, respectively. The chart compares each metric separately. It does not add them together or convert them into the accuracy of every answer you will receive. These are vendor-reported results that we have not reproduced.Tencent: Hy4 launch announcement

The direction is relevant to office work because filling an information gap with a confident claim creates review work. If meeting notes do not identify an owner, a useful action list should preserve that uncertainty. If two documents give different prices, it should retain dates and references. Reducing guessing could lower verification effort, but users still need to observe whether the model acknowledges missing evidence and remains consistent through follow-up questions.

Complex-task outcomes depend on the model, prompt, tool environment, and budget together. More retries and tool calls can improve completion while increasing time and consumption. The useful comparison unit is the cost of an accepted result, including the steps that still require a person. The public material available to us does not establish a single Hy4 win rate against all major alternatives.

5. Applications and ecosystem: choose where work happens

Tencent lists Yuanbao, ima, WorkBuddy, CodeBuddy, the WeCom assistant Dayuan, and TokenHub among the available routes, alongside integrations involving OpenRouter, Hermes, and OpenCode. This offers several ways to try the model. It does not establish identical tools, file limits, prices, or regional support across them. The sensible starting point is the type of work and the material you already have.Tencent: Hy4 launch announcement

Tencent: Hy4 launch announcement · ima: Tencent knowledge workspace · OpenRouter: Hy4 listing and regional availability

For a person whose material already lives in ima, the possible gain is less repeated uploading and less explanation of background. For someone working in a repository, the useful feature is keeping editing and verification in the same environment. These are workflow-fit judgments. We have not tested every updated application, and we do not assume that any application will permanently use Hy4 by default.

A team building its own backend should examine the model-service route rather than treating a personal tool subscription as general infrastructure. TokenHub explains that one model can have several online inference services for different business uses. That can help distinguish experimental and production consumption. Authentication, quotas, and concurrency still depend on the configured service; this review does not supply untested integration code.TokenHub: training-data policy and service overview

Third-party access adds another boundary. OpenRouter’s Hy4 page currently says EU in-region routing is unavailable. That is a specific routing limitation, not a complete list of supported countries or a promise that any account can register and pay. International users should establish their actual model access and applicable terms before committing to a tool migration.OpenRouter: Hy4 listing and regional availability

6. Privacy and control depend on the service

TokenHub’s product page states that customer data is not used for model training and links its privacy information. This is a statement about that service. It cannot automatically be extended to Yuanbao, ima, WorkBuddy, third-party routing providers, or every tool someone connects. We did not obtain a basis for giving all these routes one shared retention period, processing location, or deletion deadline.TokenHub: training-data policy and service overview

Document and coding workflows can move information through several places: a local project, the agent application, the model service, a browser session, and an exported file. A company deciding whether to provide internal material needs to understand which content leaves its environment and where results are stored. A no-training policy answers one question about use; storage, access, and sharing rules still shape the full data path.

Computer control also requires clear delegation. Information visible on a page should not, by itself, become new authorization. Preparing a draft and sending a file externally are different tasks. A suitable application should make the actual action and result understandable, with a way to stop or take over when necessary. This is an assessment of what a useful workflow requires, not an allegation of a specific Hy4 incident.

For teams considering self-hosting after release, licensing and operations remain separate decisions. The preview has a verifiable Apache 2.0 model card; the official version needs its own release material. Local inference also does not remove data flows created by external search, connectors, or application logs. Deployment location affects control, but the surrounding software determines much of the practical boundary.Tencent: Hy4-preview model card and license · Tencent: Hy4 launch announcement

7. API pricing: distinguish the promotion from the budget

Tencent Cloud lists Hy4 at CNY 6 for uncached input, CNY 18 for output, and CNY 0.3 for cached input, all per million tokens. Its current pay-as-you-go promotion charges 50% of those rates through December 31, 2026, at 23:59:59 Beijing time: CNY 3, 9, and 0.15. The personal Token Plan promotion uses a different billing arrangement.Tencent Cloud: pay-as-you-go prices and Hy4 discount

Cached input is 95% cheaper than uncached input at the listed rates: 1 − 0.3 / 6 = 95%. That comparison uses the input rate as its denominator. It applies only to the input that actually hits the cache. Output and uncached material retain their own charges. Describing this as a 95% reduction in the entire task bill would substantially overstate the saving.

Tencent Cloud: pay-as-you-go prices and Hy4 discount

These are arithmetic examples, not measured token counts. In the third row, the total list-rate bill falls 47.5% relative to the second row. That is much less than the reduction in the cached-input rate because both scenarios still contain 2M output tokens. For long reports and code generation, output length and retries deserve attention alongside caching.

TokenHub’s billing explanation says generated tokens may still be charged when a request is interrupted, including output that has not finished returning. Stopping an unhelpful task can limit further consumption without making earlier computation free. Search, conversion, execution infrastructure, and review effort also belong in a fuller budget. The examples above do not assume that all those services are included in the token price.Tencent Cloud: billing rules

Pay-as-you-go is worth considering for irregular usage, workflows requiring confirmed image support, and custom business applications. The launch discount can make a controlled evaluation cheaper. A lasting budget should also retain a list-price scenario. Otherwise, the first month may teach you more about a temporary promotion than about the economics of your workflow. We retain the original currencies rather than inventing an exchange-rate conversion.

8. Personal subscriptions: price, credits, and restrictions

The personal subscription catalog contains two families: Hy Token Plan and General Token Plan. The table gives their published monthly charges and credits per subscription month. These are tool-use plans, not a single membership for Yuanbao, ima, and every Tencent product. The available model library can change, so product access and permitted use need to be checked alongside the monthly price.Tencent Cloud: personal Token Plan subscriptions

Tencent Cloud: personal Token Plan subscriptions

The important feature restriction is easy to miss: the Hy Token Plan documentation says its available models do not currently support image or video capabilities. Someone hoping to analyze screenshots through that subscription should not rely on the image-support flag in the pay-as-you-go model catalog. General Token Plan also needs its own capability check; we do not transfer one family’s restrictions or entitlements to the other.Tencent Cloud: personal Token Plan subscriptions

A separate promotion applies to Hy4 credit deductions from November 4 through November 17, 2026, at 12:00 Beijing time. Per million tokens, uncached input, output, and cached input consume 30, 90, and 1.5 credits. At 20 credits per CNY, those correspond to CNY 1.5, 4.5, and 0.075. These are rates for consuming prepaid plan allowance, not unrestricted cash API prices for an arbitrary backend.Tencent Cloud: personal-plan promotions · Tencent Cloud: personal-plan credit deductions

The official estimates of tokens available per tier incorporate assumptions such as historical cache-hit patterns. We do not convert them into guaranteed question counts or completed projects. Conversation history, repeated file reading, and output length can materially change deductions. A favorable credit rate also does not make unused subscription spending recoverable. The relevant question is whether your permitted work will use the allowance productively.Tencent Cloud: personal-plan credit deductions

Personal-plan rules allow use in supported AI tools and prohibit custom application backends, automated scripts, and non-interactive batch calling. Unused allowance does not roll over, and the plans are non-refundable under the published rules. Those conditions are part of value for money. A frequent interactive user can examine actual credit consumption; a business backend or periodic batch workload should choose a service intended for that purpose.Tencent Cloud: personal Token Plan subscriptions

9. Usability and migration: judge an accepted task

Because this is a desk review, we do not invent click-by-click experience or a measured time to first token. We assess the access path, verification work, and cost structure instead. The following scenarios are illustrative: they describe useful acceptance criteria, not tasks we uploaded or completed. They also make the buying question more concrete than an overall claim that the model is good at office work.

For an individual, migration is usually more than registration. Knowledge-base categories, project instructions, and useful history do not automatically remain intact when changing models or products. If an existing tool already handles a job reliably, Hy4 needs to improve a recognizable bottleneck. Less material preparation or a better output format may justify a switch; a new release name alone does not cover the transition cost.

For developers, model identifiers, structured outputs, and tool calls need checking in the actual framework. Services sharing a familiar protocol can still differ in their handling of thinking messages, tool results, recovery, or very long inputs. Link observations and billing records to a specific service and version. A reproducible task is a stronger basis for a migration decision than one memorable response.

The attractions are tangible: existing applications can reduce the need to assemble a toolchain, several billing routes support different usage patterns, and a large context window creates room for substantial material. The current weaknesses are uneven capabilities between routes, unconfirmed official-release weights, and the absence of our own matched outcome data. Together, they favor a defined trial before broader adoption.

10. Hy4 versus Kimi, DeepSeek, and Claude

Compare services you can actually buy and use. We checked official information for Kimi K3.1 Pro, the current DeepSeek Flash version, and Claude Sonnet 5.5 separately. We do not treat Tencent’s competitor chart as an independent ranking. Prices retain their original currencies and billing conditions; we neither mix them with app subscriptions nor turn a token-price ratio into a claim about quality.

Tencent Cloud: pay-as-you-go prices and Hy4 discount · Kimi: official K3.1 Pro pricing and product access · DeepSeek: current model identities and official pricing · Anthropic: official Claude model pricing

Kimi distinguishes its web application, Kimi Work, Kimi Code, and API access. Users already relying on those environments should compare continuity of material and editing, not just model output. Standard K3.1 Pro and Highspeed also have separate prices. The table uses standard Pro; combining the faster version’s positioning with the standard version’s rate would describe a product that is not actually on offer.Kimi: official K3.1 Pro pricing and product access

DeepSeek’s official page identifies deepseek-flash as DeepSeek-V4.1-Flash-1017 and explains that some older Flash names route to that version. Such mappings matter for reproducing results. Its direct dollar rates also depend on peak and off-peak periods. A background job and daytime interactive work can therefore have different costs. Treat the actual model version and service as the comparison unit rather than assuming an old alias is permanent.DeepSeek: current model identities and official pricing

Sonnet 5.5 has a documented identifier and several cloud distribution routes. A team with established Claude tools and caching should ask whether a move changes the amount of checking required. Cache writes, tool execution, retries, and human acceptance can all affect total cost. A lower input rate may be attractive, but frequent rework can outweigh it on tasks where the existing combination already delivers reliable results.Anthropic: Sonnet 5.5 version and availability · Anthropic: official Claude model pricing

A practical shortlist follows the work. Tencent-heavy office users can examine Hy4’s integration; Kimi workspace users can compare continuity and delivery; direct-API buyers can include DeepSeek in a cost-sensitive evaluation; established Claude teams can measure the benefit of switching against verification effort. These can coexist in a task-based setup. We have no evidence for a blanket claim that one should replace all the others.

11. Rating and recommendation

We use six equally weighted editorial dimensions, each scored out of 10. Five means a clear use with substantial limitations; eight means good fit supported by public documentation; ten would require exceptionally strong capabilities, availability, and independent evidence. These judgments do not include an invented speed, quality, or success-rate measurement, and they are not a cross-model performance leaderboard.

The scores are task fit 8.0, product ecosystem 8.5, cost clarity 7.5, access choices 8.5, evidence maturity 6.5, and deployment consistency 5.5. The sum is 44.5; dividing by six gives 7.4167, rounded to 7.4 / 10. Established application routes and billing choices help the rating. Limited release-stage evidence, pending weights, and capability differences between routes hold it back.

Hy4 deserves a look from people with existing Tencent office material, developers prepared to judge their own accepted tasks, and application teams needing long context and a measurable token bill. Individuals can establish usefulness in an existing application before choosing a subscription. Teams can examine a specific service and the full cost of a representative task before increasing its share of work.

It is currently a poor fit for assumptions that require immediate official-release self-hosting, screenshot understanding through the personal Hy Token Plan, unrestricted backend use of personal allowance, or cross-application execution without review. These limitations concern access and evidence as much as the model itself. Understanding them makes the purchase decision more meaningful than reacting to a launch ranking.

Our conclusion is that Hy4 offers a credible route to investigate for office and coding work, especially where reducing material transfer and extending analysis into tool execution could help. Its value should be judged by editable documents, verified changes, and an acceptable bill. A candidate assigned to a clear task is easier to evaluate honestly than a new model expected to replace an entire working setup overnight.

Sources

1. Tencent: Hunyuan public debut in 2023
2. Tencent: Hy4 preview announcement
3. Tencent: Hy4 launch announcement
4. Tencent: Hy4-preview model card and license
5. Tencent Cloud: models and documented limits
6. Tencent Cloud: pay-as-you-go prices and Hy4 discount
7. Tencent Cloud: billing rules
8. Tencent Cloud: personal Token Plan subscriptions
9. Tencent Cloud: personal-plan credit deductions
10. Tencent Cloud: personal-plan promotions
11. TokenHub: training-data policy and service overview
12. OpenRouter: Hy4 listing and regional availability
13. ima: Tencent knowledge workspace
14. Kimi: official K3.1 Pro pricing and product access
15. DeepSeek: current model identities and official pricing
16. Anthropic: official Claude model pricing
17. Anthropic: Sonnet 5.5 version and availability

This review was generated with the assistance of AI technology.