# Hy4 Wins 9 of 12 Scenarios, No. 2 Rank Unproven [2026] > Source: > Published: 2026-11-06 00:17:08+00:00 Tencent released **Hy4** on November 4, 2026, an open-weight model that adds visual understanding and the ability to operate a computer. [The Decoder reported](https://the-decoder.com/tencent-releases-hy4-with-vision-and-computer-use-claims-second-place-among-chinese-ai-models/) the launch under a headline saying Tencent claims second place among Chinese AI models. That ranking is the part of the story most readers will repeat, so it deserves the closest look. Here is the short version of what holds up. Tencent’s announcement confirms specific benchmark results, including a 9-of-12 scenario lead on its WorkBuddy evaluation and a finance score of 3.04. It does not establish a universal second-place ranking across all Chinese models. We treat the second-place line as a claim, not a fact, and the rest of this analysis keeps those two categories apart. Readers who followed our earlier piece on [Hy4’s 770 billion parameters](https://shattered.io/tencent-hy4-770b-parameters-apache-2-0-open-source-model-2026/) will recognize the model size. This article focuses on what is new in the second wave of coverage: vision, computer use, the professional-software training, and the evidence behind the ranking. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Tencent Shipped on November 4 Hy4 is an open-source, open-weight model from Tencent with 770 billion parameters. Three capability groups define the release. The first is visual understanding, covering fine-grained image processing, complex chart analysis, and mathematical visual reasoning. The second is computer use through both command-line and graphical interfaces. The third is training on professional software, with Tencent naming Blender and CAD applications. TechNode summarized the launch as Tencent introducing Hy4 with vision and computer-use capabilities. That framing matches the company’s own announcement, which presents the two features as the main upgrade over earlier models in the line. Tencent did not publish attributable direct quotes from executives in the excerpts we reviewed, so this article paraphrases the announcement and names the outlets that carried it. The date matters for context. Hy4 arrived on November 4, 2026, in a market where Chinese labs ship open-weight flagships at a fast clip and each claims a lead on a chosen set of tests. Tencent follows that pattern, but the results it published center on office work and finance rather than the usual coding and math leaderboards. That choice tells you where Tencent wants Hy4 to earn its keep. An agent that can use software fits office and finance products better than a pure chat model does. We read the benchmark selection as a product signal, and we label that reading as analysis rather than as something Tencent stated. ## How Hy4 Computer Use Works Computer use means the model acts on a machine instead of only describing what to do. Tencent says Hy4 can operate both command-line interfaces and graphical user interfaces. That dual coverage matters because real work crosses both. A developer might run a script in a terminal, then check the output in a desktop app, then move a file through a file manager. ### Command line and graphical interface Command-line control is the easier half. Text in, text out, and errors come back as readable messages. Graphical control is harder because the model must read pixels, locate buttons, and decide where to click. Pairing Hy4’s visual understanding with its interface skills is the logical route to the second half, and the launch coverage treats the two features as linked. Tencent has not published, in the material we reviewed, a standalone computer-use benchmark with named competitors. The numbers it did release come from its WorkBuddy office evaluation and a finance study, covered below. Anyone planning to let Hy4 click through production systems should therefore run their own tests first. ### Blender and CAD training The most unusual detail in the announcement is the professional-software training. Tencent says it trained Hy4 for software including Blender and CAD applications. Both are complex tools with deep menus, many modes, and long task chains. A model that can drive them has to plan across dozens of steps, which is a harder test than answering a question. We do not have independent results showing how well Hy4 performs inside those programs. The claim is training intent, not measured skill. Until third-party testers publish task completion rates in Blender or a CAD package, treat the capability as promising but unproven. ## Vision Skills: Charts, Images, and Math Visual understanding is new for Hy4, and Tencent lists three strengths. Fine-grained image processing means reading small details such as labels, text, and objects in a crowded frame. Complex chart analysis means pulling structured information from graphs and tables. Mathematical visual reasoning covers problems where the answer depends on a diagram, such as geometry. Chart analysis connects directly to the finance evaluation. Analysts spend much of their day reading charts, filings, and slide decks. A model that can parse those inputs without a separate extraction step cuts a stage out of the workflow. That is a practical reason to pair vision with finance testing, and it may explain why Tencent leaned on that domain. The announcement does not give a public vision benchmark table with named rivals in the excerpts we saw. That leaves a gap. Open-weight vision models from several labs already exist, and buyers will want side-by-side results on the same chart and diagram sets. We expect independent testers to fill that gap in the coming weeks. ## WorkBuddy Results: 83.18 Average, 96.15% Pass Rate The strongest numbers in the launch come from WorkBuddy, an evaluation Tencent uses for office-style tasks. Hy4 posted an average performance score of 83.18 and a pass rate of 96.15%. It ranked first in 9 of 12 scenarios in mainland China. Those are Tencent’s figures from its own announcement, and we found no independent reproduction of them. Two cautions apply. First, the benchmark belongs to the model’s maker, which can shape task design in ways that favor its own system. Second, the 9-of-12 result applies to mainland China scenarios, so it says little about English-language or global office work. Neither point makes the result false. Both limit how far you can stretch it. The arithmetic is still worth stating. Winning 9 of 12 scenarios is 75% of the set, which leaves 3 scenarios where another model ranked higher. Tencent’s summary does not name which three, and that detail would tell buyers a lot. If the misses cluster in a single category such as spreadsheet work, a team could route that category to a different model. A 96.15% pass rate sounds high, yet pass rates depend on how strict the grader is. A lenient rubric inflates the figure. Without the rubric, the number works best as a relative signal across models graded the same way. ## The Second-Place Claim: What the Evidence Supports The Decoder’s headline says Tencent claims second place among Chinese AI models. [Silicon UK](https://www.silicon.co.uk/cloud/ai/tencent-ai-hy4-632185) ran a similar angle, saying Tencent describes its latest model as reaching the top tier. Both outlets are reporting what Tencent says about itself. Neither converts that statement into an independent ranking. Our fact check lands here: the second-place line is unconfirmed as a general claim. Tencent’s official announcement confirms benchmark results, but it does not establish a universal ranking. A universal ranking would need a shared test suite, a defined list of Chinese models, and a neutral scorer. The material we reviewed contains none of those three elements. Rankings also shift by task. A model can lead in office work, trail in coding, and sit mid-pack in math. Any single position hides that spread. When a company says second place, ask second on what, measured by whom, and against which models. Tencent’s own results answer only part of that. Independent leaderboards will settle the question over time. We covered how aggregate indexes work in our look at [Gemini 4 Carbon on the Artificial Analysis index](https://shattered.io/gemini-4-carbon-artificial-analysis-59-claude-opus-5-5-2026/), and the same process will eventually score Hy4. Until a neutral index places it, the accurate wording is that Tencent claims a top-tier result and published evidence for specific scenarios. ## Finance Blind Test: 3.04 Against Three Rivals The finance study is the closest thing in the launch to a head-to-head. In a blind evaluation, 213 finance professionals judged outputs on 579 real tasks. Hy4 scored 3.04. DeepSeek-V4.1-Flash scored 2.97, while Kimi K3.1 Pro and GLM-5.3 each scored 2.68. Blind scoring reduces the chance that graders favored a brand, which makes this result stronger than a self-run benchmark. | Model | Finance score | Gap to Hy4 | Gap as % of rival score | |---|---|---|---| | Hy4 (Tencent) | 3.04 | 0.00 | n/a | | DeepSeek-V4.1-Flash | 2.97 | 0.07 | 2.4% | | Kimi K3.1 Pro | 2.68 | 0.36 | 13.4% | | GLM-5.3 | 2.68 | 0.36 | 13.4% | The gap columns are our own arithmetic on the published scores. They show a very narrow margin over DeepSeek-V4.1-Flash and a wider one over Kimi K3.1 Pro and GLM-5.3. Scores of that size rarely mean a model wins every task. They usually mean it wins slightly more often than the others. Readers who want the Kimi side of this comparison can check our report on [Kimi K3.1 Pro going open-weight](https://shattered.io/moonshot-kimi-k3-1-pro-open-weights-2-8-trillion-parameters-2026/). That model sits at 2.8 trillion parameters, more than three times the size of Hy4, yet it trails Hy4 on this one finance study. Parameter count alone does not decide task quality, and this table is a clean example. ## Reading the Finance Gap Without Overselling It A 0.07 gap over DeepSeek-V4.1-Flash needs context. The sheet we have does not include confidence intervals, task-level variance, or the scoring scale’s maximum. Without those, a difference of 0.07 could sit inside normal noise. We would not tell a bank to switch vendors on that margin. The larger gap, 0.36 over Kimi K3.1 Pro and GLM-5.3, looks more meaningful. Even so, the study measured finance tasks only. It says nothing about legal drafting, customer support, or code generation. Teams should read it as evidence for one domain and keep their own evaluation sets for the rest. The sample size helps. With 579 tasks and 213 professionals, the study is large enough to take seriously. By contrast, many launch-day benchmarks rest on a few dozen prompts. Tencent deserves credit for choosing a design in which finance professionals, not automated scripts, graded the outputs. One more point on provenance. The finance results appear in coverage of Tencent’s announcement, and no outside lab has replicated them yet. Independent replication would raise confidence. Until then, a careful reader holds the result as likely directionally right and exact on neither the size of the lead nor its stability. ## Cybersecurity Safeguards and the Defensive-Only Training Line [Rappler reported](https://www.rappler.com/technology/tencent-hy4-ai-model-trained-limited-cybersecurity-data/) that Tencent rolled out Hy4 trained with limited cybersecurity data amid hacking fears. Tencent said its cybersecurity training stayed limited to defensive uses, such as identifying and fixing software flaws. It said it excluded material on building hacking tools or breaking into systems. [BusinessDay](https://www.businessday.co.za/world/international-companies/2026-11-04-tencent-releases-new-ai-model-pledges-safeguards-against-cyberattack-use/) carried the same story under a headline about safeguards against cyberattack use. The timing is deliberate. A model that can operate a command line and read a screen could, in the wrong hands, automate parts of an intrusion. Tencent wants regulators and customers to see that it considered that risk before release. Data filtering is a real control, but it has limits. An open-weight model can be fine-tuned by anyone who downloads it, and a determined user can add back what the original trainers removed. Filtering lowers the baseline risk without eliminating it. That caveat applies to every open-weight release, not only Hy4. Teams deploying agents with shell access should add their own guardrails regardless of the vendor’s training choices. Our walkthrough on [AI guardrails for LLM apps](https://shattered.io/ai-guardrails-llm-app-python-12-steps-2026/) covers input checks, output filters, and permission boundaries that apply to any model, including this one. ## Pricing: Reported Numbers Without Confirmation Pricing is unconfirmed in the reliable sources we reviewed. Secondary coverage reports $0.834 per million input tokens and $2.501 per million output tokens. The cited Tencent announcement does not confirm those figures. We therefore present them as reported numbers only, and we advise readers to check Tencent’s own pricing page before budgeting. We urge caution because converted prices drift often. Resellers and API aggregators round, convert currencies, and add margin, so two outlets can print different numbers for the same model on the same day. Because the weights are open, price is not the only route to Hy4. A company with enough GPUs can host the model itself. That changes the cost question from price per token to hardware, power, and staff time. For a 770 billion parameter model, the hardware bill is large, so most small teams will still call a hosted endpoint. The self-hosting economics are not hypothetical. Our report on [Revolut cutting AI coding costs with self-hosted GLM](https://shattered.io/revolut-self-hosted-glm-ai-coding-costs-10x-2026/) shows one fintech already running that math with a Chinese open-weight model. Hy4’s finance score may attract similar teams, though each will need its own cost model. ## Competitive Picture: Hy4 Next to DeepSeek, Kimi, and GLM The only direct comparison in the launch material involves three models: DeepSeek-V4.1-Flash, Kimi K3.1 Pro, and GLM-5.3. All three come from Chinese labs, all three compete for the same enterprise buyers, and Hy4 beat each of them on the finance study. That is a narrow slice of a much larger field. Notice who is missing. The sources we reviewed give no confirmed Hy4 results against Alibaba’s Qwen line or against Western frontier models. That omission is why the second-place claim stays unconfirmed. You cannot rank a model second without scores for the models above and below it. | Item | What the sources confirm | Status | |---|---|---| | Hy4 WorkBuddy score | 83.18 average, 96.15% pass rate | Tencent-reported | | WorkBuddy scenario wins | First in 9 of 12 (mainland China) | Tencent-reported | | Finance blind test | 3.04 across 213 pros and 579 tasks | Tencent-organized | | Rival finance scores | 2.97, 2.68, 2.68 | Tencent-reported | | Universal second place | No neutral ranking found | Unconfirmed | | Token pricing | Reported in secondary coverage only | Unconfirmed | For a wider view of how open-weight models stack up on specifications, our [open-weight spec sheet](https://shattered.io/open-weight-ai-models-side-by-side-the-spec-sheet/) lists the major releases side by side. Hy4 will slot into that field as a 770 billion parameter entry with vision and computer use, two features that buyers should check against each rival’s own documentation. ## Historical Context: From Chat Models to Software Operators Language models started as text predictors, moved into chat assistants, and are now becoming operators of software. Hy4 sits on that third step. Over the past two years the industry’s attention shifted from how well a model writes to how well it finishes a job inside real applications. Chinese labs have followed a distinctive path through that shift. Many released weights early and often, which let developers and rival labs study, fine-tune, and benchmark their models. That openness let developers outside China test the models directly. Hy4 continues the approach by pairing open weights with agent features. The finance focus also has precedent. Banks and brokerages were early buyers of language tools because document-heavy work suits them. A blind test with 213 professionals reflects a lesson the industry learned the hard way: self-graded benchmarks impress press releases, while expert-graded studies persuade buyers. We also see a security pattern repeating. Each time models gained new abilities, vendors paired the launch with a safety statement. Tencent’s defensive-only training pledge belongs to that tradition, and it arrives as computer use makes the stakes more concrete than they were for chat-only systems. ## Market Impact for Developers and Enterprises The immediate winners are teams that want an open-weight agent for office and finance work. Hy4 gives them a model with visual input, interface control, and a documented evaluation in their domain. The immediate pressure lands on rival labs, which now face a competitor that publishes expert-graded results instead of leaderboard screenshots. For developers, the practical question is integration. Computer use requires a sandbox, a screenshot loop, and permission controls. Open weights make that loop easier to customize, since you can inspect and adjust the model’s behavior. A hosted API gives less control but less work. Each team will pick based on compliance needs and available hardware. For enterprises outside China, the picture is more complicated. Procurement teams must weigh performance against data-residency rules, vendor risk policies, and regulator expectations. Open weights reduce some of that friction because the model can run inside the company’s own environment. They do not remove the need for a legal review. Pricing pressure is the third effect. Whatever Hy4’s final list price turns out to be, a strong open-weight option with agent features gives buyers leverage when negotiating with hosted competitors. For broader context on this beat, see our [AI and machine learning coverage](https://shattered.io/ai-machine-learning/), where we track model releases and pricing moves as they happen. ## What We Can and Cannot Say About Hy4 A short ledger helps. Confirmed by Tencent’s announcement: the November 4, 2026 release, the open-weight status, the 770 billion parameters, the vision and computer-use capabilities, the Blender and CAD training, the WorkBuddy figures, the finance study figures, and the defensive-only cybersecurity pledge. Reported but unconfirmed: the token prices. Claimed but unproven: a universal second place among Chinese models. What we cannot say is just as useful. We cannot say Hy4 beats Qwen, because no comparable score exists in the launch coverage. We cannot say it beats Western frontier models for the same reason. We cannot say how it performs inside Blender or a CAD package, since no outside test is public. We cannot say what the model costs until Tencent confirms it. Honest reporting on a launch like this means holding those lines. The Decoder, TechNode, Rappler, BusinessDay, and Silicon UK each reported the story from the company’s announcement. That agreement shows the announcement is real. It does not independently validate the claims inside it. Readers who run evaluations for their own teams can turn that ledger into a test plan. Pick five tasks from your real workload, include at least one chart-reading job and one multi-step interface job, and grade the outputs blind with two colleagues. That mirrors the structure of Tencent’s finance study at a small scale. ## Five Predictions for Hy4 and Chinese Open-Weight Models These are our expectations, not reported facts. We offer them so readers can judge our reasoning against the evidence above. - **Independent finance tests will narrow the gap.** A 0.07 margin over DeepSeek-V4.1-Flash is thin, and we expect outside testers to report a near tie on some task types. - **A neutral index will place Hy4 in the top tier, but its exact rank will vary by index.** Rankings differ by task mix, so no single position will match the second-place claim exactly. - **Computer-use benchmarks will become the next battleground.** Vision plus interface control invites head-to-head tests, and rival labs will publish their own numbers. - **Security scrutiny will grow.** Regulators and researchers will probe whether the defensive-only training holds up after[fine-tuning](https://shattered.io/category/ai-machine-learning/) , because open weights allow modification. - **Rivals will answer within months.** Alibaba, DeepSeek, Moonshot, and Zhipu all ship fast, and we already covered[Alibaba’s Qwen4 plans](https://shattered.io/qwen4-launch-alibaba-open-source-promise-doubt-470-models-2026/) . Expect agent features to appear in their next flagships. The common thread is that evidence will replace claims. Launch-day numbers get attention, but replication decides which models teams adopt. Hy4 starts with a solid set of Tencent-reported results and a well-designed finance study. Whether it holds second place or something else depends on tests nobody has run yet. ## FAQ **What is Tencent Hy4?** Tencent Hy4 is an open-weight AI model with 770 billion parameters, released on November 4, 2026. It adds visual understanding and computer-use skills, including control of command-line and graphical interfaces. **Is Hy4 really the second-best Chinese AI model?** That is unconfirmed as a general claim. Tencent’s announcement confirms specific benchmark results, but it does not establish a universal second-place ranking. The Decoder’s headline reports the claim as Tencent’s own. **What can Hy4 do with a computer?** Tencent says Hy4 can operate both command-line interfaces and graphical user interfaces. The company also says it trained the model for professional software including Blender and CAD applications. **How did Hy4 score in Tencent’s tests?** On WorkBuddy, Hy4 averaged 83.18 with a 96.15% pass rate and ranked first in 9 of 12 scenarios in mainland China. In a blind finance evaluation with 213 professionals and 579 tasks, it scored 3.04. **Which models did Hy4 beat in the finance test?** DeepSeek-V4.1-Flash scored 2.97, while Kimi K3.1 Pro and GLM-5.3 each scored 2.68. Hy4’s 3.04 led all three, though the margin over DeepSeek is small. **How much does Hy4 cost?** Pricing is unconfirmed. Secondary coverage reports $0.834 per million input tokens and $2.501 per million output tokens, but the cited Tencent announcement does not confirm those figures. **Is Hy4 safe to use for security work?** Tencent says its cybersecurity training was limited to defensive uses, such as finding and fixing software flaws, and excluded material on building hacking tools or breaking into systems. Open weights can still be fine-tuned, so deployers should add their own controls.