Tencent released Hy4 on November 4, presenting it as an upgrade to the Hy4 preview that adds visual perception and computer operation through command-line and graphical interfaces. The model is accessible through Tencent Cloud and OpenRouter, while Tencent says its weights are planned for a later release. The announcement combines multimodal input with tool use and sustained workflows; its strongest performance figures, however, remain company-reported evaluations rather than independently replicated results.
What Tencent released
Tencent’s Hy team dated its announcement “Introducing Hy4” November 4, 2026. The company describes Hy4 as a substantial scale-up in both pre-training and post-training relative to Hy4 preview, with improvements in instruction following, long-context understanding, long-horizon tasks and deep research. These are Tencent’s characterizations of its own model, not independently established comparative findings. Tencent says Hy4 “places Hy among the strongest open-source models,” although at launch the weights were not available for outside inspection or reproduction. Tencent’s announcement provides the feature descriptions and its reported test results.
Caixin’s independent report, published November 5, likewise says Tencent officially launched the model on November 4 and highlights visual understanding and computer operation. Its article characterizes Hy4 as a 770-billion-parameter model, up from a predecessor below 300 billion. That size figure is reported by Caixin, but Tencent’s publicly accessible announcement reviewed here does not give a parameter count. It should therefore be treated as an independent report rather than a primary-source specification. Caixin’s report is partly subscriber-gated, and its accessible digest contains some figures that do not map cleanly to every metric in Tencent’s detailed page.
Vision and computer use
The clearest product change is that Hy4 can accept visual information, a first for the Hy model line according to Tencent. The company says the system can identify fine-grained image details and interpret complex charts by combining visual perception with reasoning and tool use. Its page illustrates a task locating a small rooftop component in a 2.8-gigapixel aerial image. That example indicates the capability Tencent wants to demonstrate, but a showcased prompt is not a controlled measure of general accuracy.
Tencent also says Hy4 can operate both CLI and GUI environments. It describes targeted training on Blender and computer-aided design software, along with tasks that cross multiple applications. The intended step beyond answering a question is planning and carrying out a sequence of actions in software. In practice, such claims depend on the environment, available permissions, tool integrations and evaluation method. The announcement does not publish a fully reproducible protocol for measuring general computer-use reliability across third-party applications.
The release emphasizes coding and creative workflows. Tencent reports that Hy4 improved 19.1% on its GameDev evaluation and 9.8% on GameLogic compared with Hy4 preview. It says training and evaluation were developed with professional developers from Tencent Interactive Entertainment Group and that the model was adapted for Unreal Engine, Unity and Godot. These are relative improvements reported by the company; the page does not make them directly comparable to external benchmark scores or establish that users will see the same gains on arbitrary projects.
What the reported tests show
Tencent’s finance and office section gives detail about its evaluation setup. It says 213 finance professionals participated, including 153 experts from leading financial institutions, and that more than 70% had over five years of industry experience. The evaluators assessed 579 real-world financial tasks. On the company’s resulting comparison, Hy4 scored 3.04, versus 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. Tencent says that placed Hy4 first among open-weight models in this evaluation. The announcement does not provide enough methodological detail in the summary to infer broad superiority beyond this task set.
On the investment-research benchmark IRAB, Tencent reports Hy4 scored 61.63 across 101 tasks, 12.03 points higher than Hy4 preview. The page attributes the benchmark and evaluation to RabyteTech. These results record what the named organizations say they measured, but do not independently validate the benchmark’s representativeness. Tencent also says its internal financial and office evaluations reduced hallucination from 15.76% on Hy4 preview to 5.45%, and speculative errors from 23.1% to 5.6%. Those numbers belong to the company’s internal evaluation context; they should not be generalized as universal error rates.
For office software, Tencent says Hy4 can co-edit Word, Excel and PowerPoint files with people in real time, and that blind evaluations found gains in formatting, reliability and overall usability versus the preview. The announcement also reports nearly 97% task completion for document creation and editing in Manus internal evaluations. This is a vendor-published account of a partner evaluation, not a public benchmark result with a full independent protocol in the release.
In research, Tencent says Hy4 ranked first among open-weight models on TenureTrail, its internal collection of 123 research tasks across seven disciplines. It also describes use with the Hyra 2.0 harness to optimize internal pre-training infrastructure for smaller models and says Hy4 and Hyra co-developed two foundation models, HyCell and HyPDE, which Tencent plans to open-source. These claims broaden the launch story beyond a chat model, but the announcement provides no independent replication of the internal benchmark or reported R&D outcomes.
Availability, openness and pricing
At launch, Tencent listed Hy4 on Tencent Cloud and OpenRouter and in Tencent products including CodeBuddy, WorkBuddy, Yuanbao and ima. The primary page displays API prices per million tokens: $0.042 for cached input, $0.834 for input and $2.501 for output. These are the prices shown on Tencent’s announcement page; actual access, regional availability and terms may depend on the relevant product or provider.
“Open” needs a precise qualification. Tencent says it plans to release the model weights and has partnered with vLLM and SGLang communities to support the ecosystem. It says a system card covering safety evaluations and additional technical details will accompany the weight release. The launch announcement did not itself provide downloadable weights or that system card. Developers therefore could not inspect, self-host or independently test the model weights on November 4; Tencent described a forthcoming release, not weights already published.
Tencent says it restricted cybersecurity training data to defensive material related to proof-of-concept generation and vulnerability patching, and conducted red-team audits with its security teams. It reports no elevated risk in specified offensive tasks. On biosecurity, the company reports a score of 30.4 out of 100 on its Hy-BioExperimentBench and says the result remained below its high-risk threshold. These are the developer’s safety claims, not an outside audit of the full system; Tencent says further evaluation detail is intended for the system card.
How Hy4 fits the recent model cycle
The release arrives amid a fast cycle of models positioned for professional work. Recent coverage by this site has included OpenAI’s GPT-6.1 Terra, described as adding visual reasoning and computer-use features, as well as other company announcements about AI products and services. That thematic overlap is real: Hy4 and GPT-6.1 Terra both bring vision and computer interaction into the story of models doing multi-step work. It does not make the Hy4 launch a duplicate. The vendor, product line, access path, reported evaluations and future weight-release plan are different. Tencent’s release has particular relevance to the open-weight ecosystem, subject to the fact that weights were not yet released.
For buyers and developers, the distinction between capability claims and demonstrated performance is more useful than treating launch language as settled fact. Hy4’s described CLI/GUI actions, multi-application workflows and office integrations point to product ambitions. The reported task scores offer initial evidence within Tencent’s chosen evaluations. Neither establishes how consistently the model performs across external software, unfamiliar workflows or independent test sets.
What is still unknown
At publication, several practical questions remain. Tencent had not released the weights or promised system card, so outside researchers could not inspect those artifacts through the launch page. The page does not establish a universal reliability rate for computer use, nor does it show independently reproduced comparisons for the proprietary and open-weight systems named in its benchmarks. The accessible Caixin digest reports 7.2% hallucination and 5.6% guessing-error figures, while Tencent’s primary page gives 5.45% hallucination and 5.6% speculative-error figures for its internal financial and office evaluations. Because the labels and evaluation contexts differ, these values should not be collapsed into a single model-wide metric.
The defensible takeaway is narrower but significant: Tencent launched a new Hy model with company-described visual input and computer use, deployed it through named cloud and product channels, and says model weights will follow. The launch establishes availability and product direction. Whether its benchmark advantages generalize, and what independent users can verify once weights and documentation arrive, remains an open question.
FAQs
When did Tencent release Hy4?
Tencent’s primary announcement is dated November 4, 2026. Caixin’s independent report was published November 5 and says the official launch was November 4.
What is new in Hy4?
Tencent says Hy4 adds vision perception and computer use across CLI and GUI for the first time in the Hy line. It also claims stronger long-context, long-horizon, deep-research and instruction-following performance than Hy4 preview.
Can developers download or self-host Hy4 weights now?
The announcement says weights are planned for a later release, alongside a system card. It does not provide them at launch. Tencent lists availability through Tencent Cloud and OpenRouter.
Are Hy4’s benchmark claims independently verified?
The detailed scores and comparisons on Tencent’s release page are company-reported or based on named internal/partner evaluations. Caixin independently reports the launch and some details, but does not reproduce the benchmark tests.
How large is Hy4?
Caixin’s accessible report digest describes a 770-billion-parameter model, compared with less than 300 billion previously. Tencent’s primary announcement page reviewed here does not state a parameter count, so attribute that figure to Caixin rather than treating it as a primary-confirmed specification.