Tencent Hunyuan Hy4 Ships with Vision and Computer Use, But Third-Party Benchmarks Are Still Empty

On November 4th, WorkBuddy's model list gained a new entry: Hy4, tagged "limited-time free" with a multiplier of 0.00x. The details on the right note: from November 4th to 15th, free daily quotas are granted, with subscribers receiving higher allocations.

This article aims to clear up three things: what actually changed this time; how much trust to place in the official scores; and where it ranks among domestic models.

The conclusion up front. What's upgraded isn't scale, but perception and post-training. The official scores largely check out, though hallucination rate is given in two versions, and four benchmarks are labeled internal. Third-party leaderboard positions are still empty—across four boards, it hasn't appeared on any of them.

The same day Tencent issued a press release stating "Hy4's overall performance has entered the ranks of the strongest open-source models." That claim isn't excessive, but the upgrade's focus is different from before.

The base model hasn't changed; what's added is perception

770B total parameters, 49B active, MoE. Identical to the Hy4 preview open-sourced on August 28th.

The official stance is that pre-training and post-training saw "significant expansion of data and training scale." What expanded is the data fed in and the training rounds, not the model's size. After infrastructure was rebuilt in February this year, parameters climbed—Hy3 series under 300B, Hy4 series to 770B—but this version hasn't climbed further.

Officially, three capabilities are reported: 1M ultra-large context, visual understanding, and CUA computer operation. The examples given are practical: loading an entire code repository or a document of hundreds of thousands of characters at once.

Of these three, two are new; the first isn't. On the preview's model card, the context field already listed 1M, with the parenthetical "API max input 960K, max output 64K"—this version hasn't changed a single number.

First, perception. The preview had no visual understanding; images couldn't be read. Hy4 adds it for the first time—screenshots, charts, interfaces, and error messages can be read directly, and math problems with images can be solved. The official example is finding a two-meter-wide roof component in a 2.8-gigapixel aerial image—I believe that example half-way. High pixel count doesn't equal difficulty; the difficulty is locking onto a small target in dense texture, and the release didn't give recall rates.

Once vision arrived, another thing followed: it can operate a computer by itself. Officially called CUA, Computer Use Agent—it opens a browser, clicks buttons, types commands, and completes an entire workflow. It mixes command line and GUI, with specialized training for professional software like Blender and CAD. The training ground is a sandbox built by Hunyuan and Tencent Cloud's Agent Runtime team, housing hundreds of deployed applications where the model repeatedly executes tasks and verifies results. This Agent Runtime is also open to the public.

Vision and Computer Use are two sides of one coin: the model finally gets information from the interface, not just from the prompt. Neither existed before; from zero to one, with no prior baseline.

On the post-training side, changes focus on instruction following, long-document comprehension, long-horizon agent tasks, and deep research. The preview's model card itself listed two known issues: thinking too long on complex tasks, and "excessive self-verification"—repeatedly checking the same conclusion, burning tokens. At the October 29th ecosystem conference previewing the release, chain-of-thought efficiency was listed among three core upgrades—the "thinking too long" issue was officially acknowledged.

How much of an increment? Only a few verifiable numbers

Skip the flowery language in the press release; only these sets of numbers can be verified. First, the ledger from preview to formal release.

There's a small snag with hallucination rate: two different numbers appear in two official places. The Hunyuan website research page says "15.76% down to 5.45%," while the Tencent Tech Engineering WeChat article says "12.95% down to 7.2%," and media outlets split into two camps. The two numbers come from different self-built test sets, neither with complete methodology, so they can't be converted. In such cases I won't guess which is right; I look at what's consistent across both: the error rate due to guessing, which dropped from 23.1% to 5.6%, identical verbatim.

That's a three-quarters reduction. It addresses the problem of fabricating answers when information is insufficient—closer to the pitfalls seen in daily use than the two hallucination-rate figures.

The remaining sets are specific gains. GameDev rose 19.1%, GameLogic 9.8%, and XunTu Tech's IRAB financial benchmark rose 12.03 percentage points. Officials also claim WorkBuddy office scenarios score 83.18 with a 96.15% pass rate, ranking first domestically in 9 of 12 scenarios—but this set is self-measured and co-designed with Tencent's own products, so its reference value should be discounted.

How trustworthy are the official scores?

Accompanying the release was a twelve-cell evaluation dashboard. The blue bars are Hy4, with five competing models around it; the previous generation preview appears only as a dashed line within the bars, not participating in ranking—across all twelve items, that dashed line sits below the blue bar.

Eight third-place finishes, one second-place, three fourth-place, and no first-place finishes at all.

These numbers can cross-check each other. Against Claude Opus 5.5, Hy4 loses all twelve, so it necessarily has zero first-place finishes; against GLM-5.3 and DeepSeek, it wins all (GLM's OSWorld isn't given), so those two can't be among the eight third-place results; Kimi K3.1 Pro beats it six times—three in the fourth-place items, three in the third-place items. Twelve items: eight plus one plus three, add up.

The brightest single item is E-Bench 2.0 at 70.0, beating GPT-6.1 Sol's 66.8 and DeepSeek's 65.3, second only to Opus's 80.0. It's the only cell on the entire board to crack the top two.

Fairness must be acknowledged. A 49B-active model carrying a 770B footprint achieving this position is genuinely efficient. Vision and Computer Use, brought in a single version, earned an OSWorld 2.1 score of 78.9, behind only Kimi and Opus.

As for the trustworthiness of the scores themselves, only one cross-validation exists. Officials give Terminal Bench 4.0 scores of 42.9 for GLM-5.3 and 40.0 for DeepSeek 1017; the third-party LLMBoard's November 3rd issue gives GLM 41.8 and DeepSeek 42.3. The magnitude aligns, with individual differences around 2 points. At least on the public test set, officials haven't mislabeled. Cells marked internal or with asterisks should be discounted in horizontal comparisons—the asterisk means the evaluation was first conducted at release time, with no prior benchmark reference.

Comparing against domestic rivals: don't mix the two stacks

These two stacks can't be added horizontally. Tencent offers self-built evaluations and expert crowdsourcing; third parties offer general and specialized boards with different ranges, task sets, and evaluation dates. Computing a weighted average between 3.04 and 72.62 is fabrication. So they're presented separately below.

In Tencent's self-selected scenarios, the financial crowdsourcing is the most rigorous: Hunyuan partnered with Tencent Financial Technology FiT for a crowdsourced evaluation with 213 experts, 153 of them from external top financial institutions, 70% with over five years of experience, blind-testing 579 real tasks on a 4-point scale. Hy4 scores 3.04, DeepSeek-V4.1-Flash-1017 scores 2.97, and Kimi K3.1 Pro and GLM-5.3 both score 2.68. The official claim is "first among open-source models." The tasks are real business problems, and more than half the graders come from outside Tencent.

The third-party stack is separate. SuperCLUE's intelligent index from the Jul 2026 edition (Chinese evaluation, 507 tasks, six dimensions, supplemented in September) has Qwen3.8-Max-0902 at 72.62 on top, DeepSeek-V4.1-Flash at 71.81, Qwen3.8-Max 71.48, GLM-5.3 71.29, and Kimi-K3 70.68—the top five within 2 points; Tencent's only entry on the board is Hy3 at 62.13.

Shuyoumo's November domestic list ranks by match win rate (November 3rd snapshot): Kimi K3.1 Pro leads at 94.2, DeepSeek-V4.1-Flash-1017 at 88.4, Qwen3.8-Max-0902 at 86.8, GLM-5.3 at 80.0, and Hunyuan 4 Preview at 73.2, 14th domestically and #39 globally. On Vals AI's Tax Agent Bench, Hy4 preview sits 16th at 32.6%; on the same board, Qwen3.8 Max is 49.7%, DeepSeek 1017 44.4%, and GLM 39.7%.

The Hy4 official release appears on none of these four boards; it just launched today, and nobody has had time to test it. "Where does it rank" depends on which board you ask about: Tencent's own board gives it a position relative to five rivals, but independent third-party boards have nothing yet. The only things we can say now: in Tencent's chosen scenarios, it beat the three competitors on that board; on third-party general boards, the previous generation preview landed outside the domestic top ten. How much difference remains can wait until SuperCLUE and Artificial Analysis include it.

Areas not to get excited about prematurely

The most obvious shortcoming is in terminal. On Terminal Bench Science 0.1, Hy4 scores 19.8 versus Kimi K3.1 Pro's 41.7—less than half. GPT-6.1 Sol scores 60.9, Opus 5.5 scores 56.2, much further behind. On Terminal Bench 4.0, Hy4 scores 44.4 against Kimi's 66.8. When it comes to specific tasks: modifying existing code and getting it to run—fine; iteratively troubleshooting in a command line, reading error messages, adjusting the environment—not fine. If you plan to use it for ops automation or research assistance, run it yourself first; don't decide based on the press release.

Hunyuan itself acknowledged this at the end: although Hy4 is already one of the strongest open-source models, it still trails the top closed-source models by a considerable margin, and efforts will continue to expand pre-training and post-training scale to close that gap.

For those planning private deployment, one more critical point: weights haven't been released. The preview was Apache 2.0 open-sourced directly on August 28th, downloadable from Hugging Face, ModelScope, and GitCode; the official release has only said "open-source soon" without a date. If you want to run it locally, you'll have to wait.

The pricing and free-tier details should be clear. API prices haven't changed: input 6 yuan, output 18 yuan, cache hit 0.3 yuan per million tokens. The Tencent Cloud article says "input as low as 0.3 yuan"—that's the cache price, don't misread it. Limited-time tiers: TokenHub pay-as-you-go at 50% off until December 31st, Token Plan subscription credits at 2.5% off (2.5折, i.e., 25% of original price) until November 18th, and WorkBuddy/CodeBuddy daily free quotas from November 4th to 15th. Hy4 flash is still on its way, positioned as a cost-reduction option.

How I'll test it

With about ten days of free access, I plan to try two kinds of tasks. One is image-rich multi-source materials—mixing screenshots, tables, and documents to see if it truly reads them, and especially whether it fabricates when images are unclear. In the official fridge-photo demo, the model marked four unrecognizable ingredients as "to be confirmed"; I want to reproduce that myself. The other is cross-application long tasks: making it successively edit Word, Excel, and PPT, watching whether it breaks mid-way and whether state is lost.

Pure algorithm problems and terminal debugging I won't use to judge it; the two cells on the board already make that clear.

One final aside. The free quota really is small; I hit the wall just as I got used to it. The message was polite: request rate exceeded the protection threshold, normal rate limiting, with an unblock time of 18:37 that day—four hours.

Still, I don't feel I lost out. The worst part of free access is congestion; when a model goes viral, queues can make you think the interface crashed. Rate-limiting avoids congestion, so responses are always snappy.