# Tencent Hy4 vs Hy4 preview: what changed at release > Tencent Hy4 launched November 4, 2026: what changed from Hy4 preview, from vision and computer use to Tencent's benchmarks, the list price and open weights. - Canonical HTML: https://omniakey.com/blog/tencent-hy4 - Markdown: https://omniakey.com/blog/tencent-hy4.md - Published: 2026-11-04 (updated 2026-11-04) - Author: Kiki, COO at OmniaKey - Language: en - Category: Comparison - Reading time: 8 min - All posts: https://omniakey.com/blog/llms.txt --- Tencent Hy4 went live on November 4, 2026, as the official release that follows the Hy4 preview from late August. Tencent says Hy4 is the first Hy model with visual understanding and computer use (terminal and desktop apps), plus large gains in instruction following, long context and long agent tasks. The Tencent Cloud TokenHub list price is unchanged: CNY 6 input, CNY 18 output and CNY 0.3 cached input per million tokens, and Hy4 is billed at 50% of list until December 31, 2026. The release weights are not public yet; Tencent's press release says the plan is to "release model weights and system cards in the coming days". ## Tencent Hy4 vs Hy4 preview at a glance |Item|Hy4 preview|Hy4| |-|-|-| |Release|August 27–28, 2026|November 4, 2026| |Vision|Text only at launch|Images and screenshots| |Computer use|No|Terminal and GUI| |Context|1M, 64k output|1M, 64k output| |Parameters|770B, 49B active|770B (launch coverage)| |List price|6 / 18 / 0.3|6 / 18 / 0.3| |Weights|Open, Apache 2.0|Not public yet| |Model ID|`hy4-preview`|`hy4`| The preview's weights landed on Hugging Face on August 27, 2026, and Tencent Cloud's TokenHub release notes (updated November 4, 2026) list the model from August 28. The 770B total and 49B active figures are from the preview model card; The Paper put the release series at 770B on November 4, 2026, and no Hy4 model card is public yet. Price cells are CNY per million tokens (input / output / cached input); context and output come from Tencent Cloud's model list. ## What changed, in Tencent's words Tencent's November 4 press release names three changes: vision, computer use and better work on long tasks. Vision is the headline. "The new model also introduces visual understanding capabilities for the first time," the release says, listing fine-grained image processing, complex chart analysis and mathematical visual reasoning among the new tasks. A hands-on test by the Chinese tech outlet Zhidx (智东西) the same day shows the everyday version: screenshot the part of an interface you want changed and point at it. Computer use is the second addition: "Hy4 introduces computer-use capabilities for the first time, enabling it to interact with both command-line interfaces and graphical user interfaces." Tencent also trained it further on professional desktop software, naming Blender and CAD, which it says sharpened planning on long, multi-step jobs. The third is broader: better instruction following, long-context comprehension and long-horizon agentic work. Tencent's Chinese release blog, as quoted by Zhidx, says Hy4 markedly expanded pre-training and post-training scale. Tencent has not said whether the API accepts video or PDF files; the press release mentions images only. ## Tencent's benchmark numbers for Hy4 Every figure below is from Tencent's November 4 press release, and every one is Tencent's own test or a test it arranged. The finance test is a blind evaluation on 579 real-world tasks, scored by 213 financial professionals, 153 from outside institutions and about 70% with more than five years in the industry; Tencent's Hy team ran it with Tencent Financial Technology. |Model|Score| |-|-| |Hy4|3.04| |DeepSeek-V4.1-Flash-1017|2.97| |Kimi K3.1 Pro|2.68| |GLM-5.3|2.68| The [DeepSeek-V4.1-Flash-1017](https://omniakey.com/blog/deepseek-v4-1-flash-1017-review) score of 2.97 is Tencent's measurement, not DeepSeek's own; [Kimi K3.1 Pro](https://omniakey.com/blog/kimi-k3-1-pro-review) and GLM-5.3 both scored 2.68 in the same test. Tencent's other release-day numbers: |Test|Hy4 result| |-|-| |iRaB, vs Hy4 preview|+12.03 points| |WorkBuddy average|83.18| |WorkBuddy pass rate|96.15%| |Hallucination rate|12.95% to 7.2%| |Unsupported-assumption errors|23.1% to 5.6%| |GameDev, vs preview|+19.1%| |GameLogic, vs preview|+9.8%| iRaB is the internal investment-research agent benchmark of Rabyte, an outside financial-AI company, and 12.03 is percentage points over the preview. The WorkBuddy average and pass rate come from an evaluation on Tencent's WorkBuddy, where Hy4 ranked first in 9 of 12 scenarios in the mainland Chinese market. The hallucination and assumption figures are Tencent's internal evaluations. Independent numbers exist only for the preview. Vals AI's Vals Index, dated October 30, 2026, scores Hy4 preview at 53.21%, 22nd overall and 9th among open models; LHTB 2.0, a long-horizon terminal benchmark, has the preview solving 15 of 77 tasks at pass\@0.95, a result its maintainers added on October 11, 2026; Artificial Analysis had no Hy4 entry on November 4, 2026. We have not run Hy4 ourselves. ## Where Hy4 still falls short Tencent says so itself. Its release blog, as quoted by Zhidx on November 4, 2026, says Hy4 still has a fairly large gap to the top closed models on Tencent's hard sets for frontier maths, long-horizon development, model iteration and visual QA, and is still some distance behind GPT-6.1 Sol and Claude Opus 5.5. The English press release adds that open-ended experimental design is "more limited compared to frontier models". The preview's two known issues are the ones to watch: reasoning longer than necessary on complex tasks, and a tendency to over-verify its own work (preview model card). On small jobs the opposite showed up: in Zhidx's November 4 test a small code edit took about one to two minutes, with a shorter chain of thought than the preview's. Computer use works but is slow. In the same test Hy4 operated WeChat and posted comments end to end, with precise targeting but execution "still fairly slow". ## Tencent Hy4 price and API access Hy4 costs what the preview costs. Tencent Cloud's TokenHub price table, updated November 4, 2026, lists CNY 6 per million input tokens, CNY 18 per million output tokens and CNY 0.3 per million cached input tokens for both models, and bills Hy4 at 50% of that list price from publication until December 31, 2026, 23:59:59 Beijing time. The discount note names Hy4 only. |Model|Input|Output|Cached input| |-|-|-|-| |Hy4 (`hy4`)|6|18|0.3| |Hy4 preview (`hy4-preview`)|6|18|0.3| Call it with the model ID `hy4` and TokenHub's OpenAI-compatible base URL, `https://tokenhub.tencentmaas.com/v1`. Tencent's Hy4 guide lists three protocols for the same model: OpenAI Chat Completions, OpenAI Responses and Anthropic Messages. The context window is 1M tokens, the maximum input 960k and the maximum output 64k, per the TokenHub model list. > The base URL already ends in > > `/v1` > > , so a chat request adds only > > `/chat/completions` > > . Tencent's launch post says Hy4 is live in WorkBuddy, CodeBuddy, Yuanbao, the WeCom assistant Dayuan, ima and Tencent Cloud TokenHub, and names OpenRouter, Hermes and OpenCode as API platforms that carry it. OpenRouter's model list on November 4, 2026 shows it as `tencent/hy4`, created that day, and still lists `tencent/hy4-preview` separately. ## Open weights: what is public today Only the preview is downloadable today. Its weights are on Hugging Face, ModelScope, GitCode and CNB under Apache 2.0, in BF16 with a separate FP8 checkpoint, and the card points to prebuilt vLLM and SGLang images. The card gives 770B total parameters, 49B active per token, a 1M-token context and a vocabulary of 120,832. The release is not. When we checked on November 4, 2026, Tencent's Hugging Face account listed only the two preview repositories (base and FP8), and the Tencent-Hunyuan GitHub organization had no Hy4 repository. Tencent's press release says it plans to release the weights and system cards "in the coming days"; the licence for the release is not stated yet. Size is the other barrier: a Hugging Face user wrote on August 29, 2026 that "having 30-200B model will help else people will just look and leave" ([discussion #10](https://huggingface.co/tencent/Hy4-preview/discussions/10)). ## What is still unknown - When the Hy4 weights arrive, and under which licence. - A computer-use benchmark score. Tencent has published none on any page we could read. - Whether the API accepts video or PDF files. - Hy4 Flash: Tencent's launch post says a cheaper Hy4 flash will be released soon, with no date, model ID or price. - Hy5: Tencent Tech (腾讯科技) reported on November 4, 2026 only that the next generation will be larger, attributed to unnamed sources. ## Frequently asked questions ### Is Tencent Hy4 open source? Tencent says it will be: its November 4 press release plans the Hy4 weights and system cards "in the coming days", but no public Hy4 repository existed on Hugging Face or GitHub that day. The preview weights are already public under Apache 2.0. ### How much does the Tencent Hy4 API cost? The TokenHub list price is CNY 6 per million input tokens, CNY 18 output and CNY 0.3 cached input, and Hy4 bills at 50% of list until December 31, 2026, 23:59:59 Beijing time. The preview is the same. ### Can I run Tencent Hy4 locally? Not today: the Hy4 release weights are not public. The preview weights are: 770B total parameters with 49B active, in BF16 and FP8, with prebuilt vLLM and SGLang images. ### What is the difference between Hy4 and Hy4 preview? Hy4 is the November 4, 2026 release; Hy4 preview is the August 2026 model it replaces. Hy4 adds visual understanding and computer use, with Tencent-reported gains in instruction following, long context and long agent tasks. The list price, the 1M-token context and the 64k output limit are the same; only the preview's weights are public. ## Sources checked - [Tencent: Tencent launches Hy4](https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/) - [Tencent Cloud TokenHub: Hy4 API guide](https://cloud.tencent.com/document/product/1823/132252) - [Tencent Cloud TokenHub: release notes](https://cloud.tencent.com/document/product/1823/130675) - [Tencent Cloud TokenHub: pricing](https://cloud.tencent.com/document/product/1823/130055) - [Tencent Cloud TokenHub: model list](https://cloud.tencent.com/document/product/1823/130051) - [Tencent launch post, November 4, 2026](https://news.qq.com/rain/a/20261104A074Z600) - [Zhidx (智东西) hands-on test, November 4, 2026](https://news.qq.com/rain/a/20261104A08NEM00) - [Tencent Tech (腾讯科技) on Hy4 Flash and Hy5, November 4, 2026](https://news.qq.com/rain/a/20261104A06OU500) - [The Paper on the Hy4 series at 770B, November 4, 2026](https://www.thepaper.cn/newsDetail_forward_34403267) - [Hugging Face: tencent/Hy4-preview model card](https://huggingface.co/tencent/Hy4-preview) - [Hugging Face: Tencent's models](https://huggingface.co/tencent) - [GitHub: Tencent-Hunyuan](https://github.com/Tencent-Hunyuan) - [OpenRouter: tencent/hy4](https://openrouter.ai/tencent/hy4) - [Vals AI: Vals Index](https://www.vals.ai/home) - [LHTB 2.0 benchmark](https://github.com/zli12321/LHTB) - [Artificial Analysis Intelligence Index](https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index) - [Hugging Face discussion #10](https://huggingface.co/tencent/Hy4-preview/discussions/10) Updated November 4, 2026. We will revise the Hy4 status when Tencent publishes the Hy4 release weights or Hy4 Flash.