Tencent Hy4 Is Out: What Is Confirmed and What Is Not Yet

On November 4, 2026, Tencent's Hy team published "Introducing Hy4" on its research site in Chinese and English. The post says Hy4 is the full release that follows the Hy4 preview of late August, adds vision and computer use for the first time, and is available on Tencent Cloud and OpenRouter at the same API price as the preview. OpenRouter's catalogue listed the route the same day. The weights are not public: Tencent says it plans to open-source them and will publish a system card when it does.

That last sentence is why this post exists. Marketplace listings, vendor posts and model cards each confirm different things, and a reader who takes any one of them for the whole picture will get a detail wrong. The listing says "official release" and links to a weights page that does not resolve. The vendor post says "one of the strongest open-weight models available today" about a model whose weights it has not released. The only open-weight Hy4 artifact anyone can download, as of November 5, 2026, is the August preview.

Every figure below comes from Tencent's release post, the Hy4 preview model card on Hugging Face, or OpenRouter's catalogue and model page, all read on November 5, 2026. Benchmark results in Tencent's post are published as images we could not read as text, so the numbers quoted here are the ones Tencent wrote in prose, and each is labelled as Tencent's own.

A model release in 2026 arrives through at least three documents with different authors and different reliability. The vendor's post is the authoritative statement of what shipped and at what price. The marketplace listing is a serving provider's record of what it can route, with text usually supplied by the vendor and dates set by the listing system. The model card on Hugging Face is the only one that proves weights exist, because it sits next to the files.

For Hy4 the three line up on some points and diverge on one. Tencent's post and OpenRouter's listing agree on the date, November 4, 2026, on vision input and on the list price; the 1M context appears in OpenRouter's listing, not in Tencent's post. They diverge on weights. OpenRouter's model page shows a "Model weights" link to huggingface.co/tencent/Hy4; that address returned HTTP 401 on November 5, and Tencent's post says the weights are planned. The Hugging Face card that does exist is for the preview, created August 27, 2026, and it describes a different, earlier model.

An example of the error this prevents: a post that reads OpenRouter's "Hugging Face id: tencent/Hy4" field and writes that Hy4's weights are on Hugging Face would be wrong on November 5. The failure case runs the other way too: a reader who only saw the listing and found no Tencent page might conclude the release was unconfirmed, when Tencent's own post has been live since the morning of November 4.

The post describes Hy4 as a substantial scale-up of both pretraining and post-training over the preview, with vision perception and computer use added for the first time. Computer use here means working across command-line and graphical interfaces, with targeted training on Blender, CAD software and workflows that span several applications. The post also claims gains in instruction following, long-context understanding, long-horizon tasks and deep research, and says a faster, cheaper Hy4 flash is planned.

The evidence Tencent offers is internal. In a finance evaluation run with Tencent Financial Technology, 213 professionals, 153 of them from outside institutions, scored 579 tasks; Tencent reports Hy4 at 3.04 against DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68 and GLM-5.3 at 2.68, and calls that first among open-weight models. It reports a hallucination rate falling from 15.76% on the preview to 5.45%, speculative errors from 23.1% to 5.6%, and improvements of 19.1% on GameDev and 9.8% on GameLogic, all on Tencent's own evaluation sets. A benchmark chart against other models is published as an image whose values we could not read.

Four product testimonials follow, from Manus, the financial AI agent company RabyteTech, and Tencent's own WorkBuddy and Yuanbao teams. Manus reports a document task completion rate of nearly 97%; RabyteTech reports 61.63 on its IRAB benchmark, up 12.03 from the preview; Yuanbao reports 72.53 against 65.25. These are quotes Tencent chose to publish, and two of the four come from Tencent products. The failure case is reading any of them as a third-party benchmark. The example of the right reading is simpler: Tencent is telling you which workloads it optimised for, which is useful information about where to test first.

Tencent's English post lists Hy4 at $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input tokens, and says the model ships at the same API price as the preview. The Chinese post lists ¥6, ¥18 and ¥0.3. Both are vendor-published regional list prices; neither is a conversion of the other, and we do not convert them here. The $0.834 figure matches the Hy4 preview's OpenRouter route price, which our OpenRouter price-change study tracked as a route through October.

OpenRouter's model page shows the single provider, Tencent Cloud, at $0.834 and $2.501 struck through and replaced by $0.417 and $1.2505, with $0.042 cached input halved to $0.021, under a "50% off" label. The catalogue API returns the discounted figures as the live prices: $0.000000417 per input token, $0.0000012505 per output token. No end date for the promotion appears on either OpenRouter surface or in Tencent's post. The route's context is 1,048,576 tokens with a maximum completion of 65,536, the model accepts text and image input, and the catalogue records a Hugging Face ID of tencent/Hy4 that does not resolve.

An example of why the distinction matters: a team forecasting a year of spend on the promotional rate would under-budget by half if the promotion ends and the list price applies. The failure case is printing $0.417 as Hy4's price. It is a promotional route price on one marketplace on November 5, 2026; Tencent's own stated price is $0.834.

Tencent's wording is precise and worth keeping precise. The English post says, "We plan to open-source Hy4's model weights", names the vLLM and SGLang communities as ecosystem partners, and adds that a comprehensive system card with safety evaluations will be published when the weights are released. A later section of the same post calls Hy4 "one of the strongest open-weight models available today". Both statements are in the document; only the first is about the present.

We checked the places weights would appear. The Hugging Face page at tencent/Hy4 returned 401 to both a page request and the API, the GitHub repository Tencent-Hunyuan/Hy4 returned 404, and Tencent's GitHub organisation listed Hy4-preview but no Hy4. The preview repository, created on Hugging Face at 08:52 UTC on August 27, 2026 and last modified August 28, remains the only Hy4-family checkpoint anyone can download. Its card describes a mixture of experts with 770B total and 49B active parameters, 78 layers, 256 routed experts with 8 active per token plus one shared expert, a 1M context and an Apache 2.0 licence.

Those preview figures are the ones that will get misattributed. The final Hy4 has no published parameter count, no published licence and no card, and Tencent's post describes a substantially scaled-up pretraining run, so the final model may not share the preview's size at all. The example of the right phrasing: "Hy4 preview is a 770B-parameter Apache 2.0 model; Hy4's final weights, size and licence are unpublished." The failure case is any sentence that puts "Hy4" and "Apache 2.0" together without the word preview. Our announcement-to-weights gap ledger tracks exactly this interval across vendors; Hy4 is a case it does not list yet.

The preview card, published in late August, was candid about its limits. It called itself an early version with real headroom left in pretraining and post-training, and named two known issues: spending longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work. It also published one comparative result, a blind evaluation by 163 internal experts on 203 engineering tasks in which the preview scored 2.99 against GLM 5.3's 2.92 and Kimi K3's 2.94.

The release post does not revisit those two issues by name. What it does claim is the list in section 02: vision, computer use, better instruction following, longer-horizon work, lower hallucination, and co-editing of Word, Excel and PowerPoint files in real time. Reasoning effort is still exposed; OpenRouter's model page carries a reasoning-effort setting, and the preview card documented a default of high with a no-think option. Our earlier Hy3 launch coverage and Hy3 routing guide describe the generation this one replaces.

An example of how to use the comparison: a team that tried the preview in September and found it slow on complex tasks has a specific thing to re-test, because Tencent has not said whether over-reasoning was fixed. The failure case is assuming "major gains" covers it.

Tencent's post includes a section on responsible release. It says cybersecurity training data was limited to defensive material around proof-of-concept generation and vulnerability patching, with exploit development, reverse engineering and penetration-testing content filtered out. It reports red-team audits by Tencent's Xuanwu and Keen Security Labs covering cryptanalysis, reverse engineering and exploitation, with the conclusion that Hy4 shows no elevated risk on those offensive tasks.

On biology, the post gives one number: 30.4 out of 100 on Hy-BioExperimentBench, an internal measure of open-ended experimental design, described as well below Tencent's high-risk threshold. Internal experts audited reasoning traces and found, in Tencent's words, no evidence that the model meaningfully lowers the barrier to biological harm. All of this is self-assessment by Tencent's own labs against Tencent's own thresholds; the post itself says the fuller evaluation arrives with the system card.

That timing is the point for an enterprise buyer. A system card is promised for the day the weights ship, and the weights have no date. An example of the practical consequence: a procurement process that requires a vendor safety document before an API model can be used has nothing to file yet beyond this post. The failure case is treating the section as a system card; it is a summary of one that does not exist yet.

The practical position is clear once the three sources are kept apart. Hy4 is a hosted model you can call today at a known list price and a temporary discount, with a 1M context and image input. It is not yet a model you can download, licence or audit. Teams that route work by capability and price, the pattern our OpenRouter token-share analysis shows is already common for Chinese models, can test it now. Teams that need weights should keep using the preview and watch the Hugging Face endpoint.

For organisations deciding how open-weight and hosted Chinese models fit a stack, including the procurement and data-residency questions a Tencent Cloud dependency raises, our AI transformation practice works through those choices with the vendor's documents in hand.