Tencent's new flagship can open Blender, build a scene, read a screenshot of a spreadsheet, and click its way through a three-step task in software it has never seen — per Tencent's launch demo. It can work a command line and a graphical interface in the same session.
What it cannot do, as of this week, is leave the building.
Hy4 (混元 Hy4) went out to the public on November 4, 2026, as an API product: Tencent Cloud's TokenHub, OpenRouter, WorkBuddy, CodeBuddy, and the company's Yuanbao (元宝) and ima apps. The weights are not in the release. Tencent says they and a safety report will follow, without giving a date.
What changed from the preview
The August 28, 2026 preview was text-only. Hy4 is the first model in Tencent's Hy line to ship with image understanding and computer use (电脑操作) in the same checkpoint.
The capability list is specific rather than generic, which is usually a good sign:
- Screen reading: charts, documents, screenshots, several images compared at once.
- Graphical interface control: moving through software the way a person does — clicking, typing, navigating menus — alongside command-line execution.
- A one-million-token context, with a 960,000-token input ceiling and 64,000-token output limit.
- Tool calling and vision in the same request path.
Third-party trackers list the model at roughly 770 billion parameters in a mixture-of-experts (混合专家) design. Tencent has not published a full architecture note, so treat that figure as provisional rather than official.
Trained on office work, not conversation
The training emphasis is the most interesting thing about Hy4, because it is so narrow.
Tencent says it concentrated on office work, financial analysis and game development — including design tools such as Blender and CAD, and the Unreal, Unity and Godot engines, with the company's own game studios preparing test material. That is a deliberate bet that the valuable agent work of the next two years looks like a person fighting a spreadsheet and a 3D viewport, not a person chatting.
One number illustrates the approach. In a test of 579 real-world finance tasks, judged with finance professionals, Tencent reports Hy4 scoring 3.04 against 2.97 for a DeepSeek model. A 0.07 margin on a bespoke rubric is not a landslide, but it is the kind of narrow, domain-specific measurement that says more about intent than a headline benchmark.
Tencent also reports that hallucination dropped sharply from the preview: fabricated facts in 5.45% of answers, down from 15.76%. That is still roughly one invented fact in twenty, which is a number worth remembering the next time someone describes an agent as reliable.
The scores, and who produced them
Hy4's reported results put it in the conversation with the closed frontier without ever quite beating it:
- OSWorld 2.1 (computer use): 78.9, against Claude Opus 5.5 at 81.8 and Kimi K3.1 Pro at 79.6. Tencent's own chart shows it ahead of GPT-6.1 Sol at 75.2.
- Terminal-Bench 4.0: 44.4 as reported by Tencent. An independent leaderboard run by Snorkel AI, using the Claude Code harness, lists Hy4 at 39.1 — the same model and task set, a different harness and runner.
- APEX-Agents v1.1 (office agent tasks): 75.2, ahead of DeepSeek V4.1 Flash (72.9), Kimi K3.1 Pro (68.2) and Claude Opus 5.5 (74.8).
- DeepSWE v1.1 (coding): 74.9.
- Humanity's Last Exam with tools: 59.1, behind Kimi (61.8) and Opus 5.5 (66.2).
The pattern holds across the table: strong on structured office and computer-use work, still behind on terminal-heavy coding and hard research maths. Tencent itself concedes a large gap to the top closed models on research-grade mathematics, long-horizon coding, self-improvement and advanced visual question answering. Companies rarely volunteer that list, and it is worth quoting when the marketing deck arrives.
The price is the argument
At RMB 6 per million input tokens and RMB 18 per million output tokens (≈US$0.83 / US$2.50, or HK$6.5 / HK$19.6), with cached input at RMB 0.3 per million tokens, Hy4 sits in a different pricing universe from the top of the American price list. TokenHub is discounting eligible usage by 50% through December 31, 2026, and WorkBuddy and CodeBuddy carried a free allowance from November 4 to 15.
That gap does real work. If a computer-use agent needs to look at a screen two hundred times to finish one task, the model that costs a fraction as much per look gets tried first — even if it is three points behind on OSWorld.
The word "open" is doing heavy lifting
Here is the awkward part of the launch.
The August preview shipped as open weights (开源权重) under Apache 2.0. Anyone could download it. The finished, more capable Hy4 does not, at least not yet. The GitHub page listed the release as forthcoming; there was no download link, no licence, and no system card on day one. On Hugging Face, the only Hy4 files available were the August preview.
Meanwhile, in the same fortnight, Moonshot published a 2.8-trillion-parameter Kimi K3.1 Pro checkpoint, Baidu released an ERNIE 5.5 Flash Base under a custom, region-restricted licence, and Meituan released a 1.6-trillion-parameter model under MIT. Tencent's claim that Hy4 ranks among the strongest open models is a statement about a plan, not a file.
That distinction matters for anyone making a procurement decision. An API product can be repriced, region-locked or withdrawn. A downloaded checkpoint cannot. Until the weights exist, Hy4 is a service with an unusually capable demo.
Honest limitations
- Every benchmark figure above except the Snorkel figure is Tencent's own and has not been independently reproduced. The OSWorld comparison against Opus 5.5 and Kimi uses vendor-reported numbers on all sides.
- The 770-billion parameter count is third-party, and Tencent has not published active-parameter or architecture details.
- The finance-task study used a bespoke rubric with professional judges; a 0.07 margin over another model is within the range where rubric design matters more than model quality.
- Computer use is not the same as safe computer use. An agent that can click through your software can also click the wrong button, and nothing in a benchmark score captures that.
- No weights means no independent inspection of training data, safety tuning, or the company's own claims. This article did not test the model directly.
What readers can do now
- Benchmark it on one repetitive screen task you already hate. A form-filling loop, a report export, a weekly reconciliation. Hy4's free allowance has ended, but at US$0.83 per million input tokens the test costs less than the coffee you will drink while it runs.
- Do not plan a private deployment around Hy4 yet. If your architecture assumes downloadable weights, this is not that model this month. Design the fallback now rather than at contract time.
- Read the concession, not the chart. Tencent's own list of weak spots — research maths, long-horizon coding, self-improvement, advanced visual QA — is a better routing guide than any single benchmark row.
