# Tencent's Hy4 is free in WorkBuddy from November 4 to 15

> Original title: WorkBuddy 里 Hy4 限免：11.04 - 11.15

- Source: meng shao
- Published: 2026-11-04T13:05:42.000Z
- Bad Influence: https://badinfluence.news/items/aojozrxhnuy2dx27hctamxhod
- Original: https://x.com/shao__meng/status/2117966517954572336

## Summary

Tencent's Hy4 model is free to use in WorkBuddy from November 4 to 15, according to the author's post and screenshot. The quoted Tencent announcement describes Hy4 as the first Hunyuan model with vision and Computer Use, with weights promised but not yet released.

## Why it matters

The post shows where users can access Hy4 free in WorkBuddy and for how long, which matters for anyone deciding whether to try it.

## Post

WorkBuddy 里 Hy4 限免：11.04 - 11.15

用 WorkBuddy 的朋友们，可以蹬起来了！

![](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FHWSHNesbIAAM0ca.jpg&mode=thumb&exp=1794009600&sig=cd8d23bb87acde92)

## Quoted post @shao__meng

> 腾讯混元正式发布 Hy4：首次能看图、能操作电脑，开源权重还在路上，团队对 Hy4 与顶级闭源模型的差异非常坦诚！
> 
> Hy4 第一次加上视觉和 Computer Use，主打金融、办公、前端游戏和科研。已进入最强开源阵营，但团队也坦诚和顶尖闭源模型仍有明显差距；权重「计划开源」，眼下主要通过 API 和腾讯产品可用。
> 
> https://hy.tencent.ai/research/hy4
> 
> # 模型信息
> 
> Hy4 总参数约 770B，MoE 架构，上下文 1M token。相对 8 月的 Hy4 preview，正式版大幅加了预训练和后训练。新能力有两块：第一次能看图，能做细粒度识别、复杂图表分析和视觉数学推理；第一次支持 Computer Use，命令行和图形界面都能操作，还专门练过 Blender、CAD 这类专业软件，以及跨应用的长周期任务。
> 
> # 几个主打场景
> 
> 金融方面，腾讯和 FiT 做了几千个工作区、上万个任务，覆盖基金研究、个股分析、量化、资产配置、公司财务和税务。213 位金融从业者参与评测，其中 153 人来自外部机构，超过七成有五年以上经验，共 579 个真实任务。按四分制，Hy4 得 3.04，DeepSeek-V4.1-Flash-1017 是 2.97，Kimi K3.1 Pro 和 GLM-5.3 都是 2.68。搜索相关的幻觉率从预览版的 15.76% 降到 5.45%，靠猜测出错的比例从 23.1% 降到 5.6%。对 Word、Excel、PowerPoint 的协同编辑也做了加强；合作方 Manus 称办公文档任务完成率接近 97%，这是对方自报数字。
> 
> 办公和协同方面，在 WorkBuddy 评测里平均分 83.18，任务通过率约 96%。接上企业微信后，它可以把聊天、邮件、文档、日程和会议串起来，整理多项目进度、待办和优先级。
> 
> 前端和游戏方面，加强了复杂交互、业务逻辑和状态管理，并加深了对 UE、Unity、Godot 的适配。相对预览版，GameDev 和 GameLogic 分别提升约 19.1% 和 9.8%，这是和自己比，不是和别家比。
> 
> 科研方面，内部基准 TenureTrail 有七个学科、123 个可执行研究任务，Hy4 在开源权重模型里排第一，在全部对比模型里大约是第三（约 53.6，Claude 约 59）。它在数学构造、物理建模、地球科学、系统工程和社会经济决策上表现较好。一个官方案例是：给 84 张无序、部分镜像的二维切片，含约 699 万个细胞，Hy4 在离线、没有专用配准工具的情况下，约 1.5 小时完成小鼠胚胎三维重建。
> 
> # AI 研发与开源副产品
> 
> 混元在为「递归自我改进」做准备。Hy4 配上 Hyra 2.0 开发框架，能优化同架构小模型的预训练基础设施，并自动训出在多项评测上超过官方 Jev 基线的变体。还开源了两个基础模型：做单细胞表征的 HyCell，以及做物理场建模的 HyPDE。
> 
> # 公开榜单表现如何
> 
> 腾讯自己跑了一组对比。在非内部的公开类评测里，Hy4 多数排在第二到第四；十二项主表里没有总榜第一，最好成绩之一是 E-Bench 2.0 的第二。开源阵营里和 Kimi K3.1 Pro 大致打平，各有胜负：代码、三维建模、检索和科研偏 Hy4，终端自动化和科学终端任务偏 Kimi。
> 
> Agent 评测 APEX-Agents v1.1 上，腾讯表里 Hy4 是 75.2，略高于它测到的 Claude Opus 5.5（约 74.8）和 DeepSeek（约 72.9）。但在其他多数行上，Opus 仍领先。腾讯自己也写了：和顶尖闭源模型仍有明显差距，尤其在前沿数学、超长开发任务、模型迭代和视觉问答上。例如 Hy-FrontierMath T5 大约只有 13，和 GPT 系接近满分差得很远。
> 
> # 价格、入口与安全
> 
> 标价和预览版一样：输入每百万 token 6 元，输出 18 元，缓存命中 0.3 元。腾讯云 TokenHub 优惠到 2026-12-31，打五折。WorkBuddy / CodeBuddy 在 11 月 4 日到 15 日有一定免费额度。已上线 TokenHub、OpenRouter、WorkBuddy/CodeBuddy、元宝、企业微信「大元」、ima 等；权重「计划开源」，并会配系统卡，目前正式版权重还没放出来。更便宜的 Hy4 flash 也说快来了。
> 
> 安全上，网络安全训练只保留防御向内容，不含利用、逆向和渗透；生物安全自测 Hy-BioExperimentBench 约 30.4/100，团队称未发现明显降低实施生物危害的门槛。

https://x.com/shao__meng/status/2117934508985196730

### Quoted post, AI translation

> Tencent Hunyuan has officially released Hy4: the first version that can see images and operate a computer. Open weights are still on the way, and the team has been candid about how Hy4 differs from the top closed-source models!
> 
> Hy4 adds vision and Computer Use for the first time, focusing on finance, office work, frontend and game development, and scientific research. It has entered the strongest open-source tier, but the team also openly acknowledges a clear gap with the leading closed-source models. The weights are "planned to be open-sourced"; for now, it is mainly available through the API and Tencent products.
> 
> https://hy.tencent.ai/research/hy4
> 
> # Model Information
> 
> Hy4 has about 770B total parameters, uses a MoE architecture, and supports a 1M-token context. Compared with the Hy4 preview from August, the official version has received substantially more pretraining and post-training. There are two new capability areas: first, it can now view images, performing fine-grained recognition, complex chart analysis, and visual mathematical reasoning; second, it supports Computer Use for the first time, operating via both the command line and graphical interfaces. It has also been specifically trained on professional software such as Blender and CAD, as well as long-horizon tasks spanning multiple applications.
> 
> # Key Featured Scenarios
> 
> In finance, Tencent and FiT built thousands of workspaces and tens of thousands of tasks, covering fund research, stock analysis, quantitative work, asset allocation, corporate finance, and tax. 213 financial practitioners took part in the evaluation, 153 of them from external institutions, with over 70% having more than five years of experience, across 579 real-world tasks. On a four-point scale, Hy4 scored 3.04, DeepSeek-V4.1-Flash-1017 scored 2.97, and Kimi K3.1 Pro and GLM-5.3 both scored 2.68. The hallucination rate for search-related tasks dropped from 23.1%... wait, correction: the hallucination rate for search-related tasks fell from 15.76% in the preview to 5.45%, and the proportion of errors caused by guessing fell from 23.1% to 5.6%. Collaborative editing for Word, Excel, and PowerPoint was also strengthened; partner Manus says its office document task completion rate is close to 97%, though this is a self-reported figure from the partner.
> 
> In office and collaboration work, it averaged 83.18 on the WorkBuddy evaluation, with a task pass rate of about 96%. When connected to WeCom (Enterprise WeChat), it can link chats, emails, documents, calendars, and meetings together, organizing multi-project progress, to-do lists, and priorities.
> 
> In frontend and games, complex interaction, business logic, and state management were strengthened, and adaptation to UE, Unity, and Godot was deepened. Compared with the preview, GameDev and GameLogic improved by about 19.1% and 9.8% respectively; this is a comparison against its own earlier version, not against other companies' models.
> 
> In scientific research, on TenureTrail, an internal benchmark with seven disciplines and 123 executable research tasks, Hy4 ranks first among open-weight models and roughly third among all compared models (about 53.6, versus about 59 for Claude). It performs well in mathematical construction, physical modeling, earth sciences, systems engineering, and socioeconomic decision-making. One official case: given 84 unordered, partly mirrored 2D slices containing about 6.99 million cells, Hy4 completed a 3D reconstruction of a mouse embryo in about 1.5 hours, offline and without dedicated registration tools.
> 
> # AI R&D and Open-Source By-Products
> 
> Hunyuan is preparing for "recursive self-improvement." Paired with the Hyra 2.0 development framework, Hy4 can optimize the pretraining infrastructure for small models of the same architecture and automatically train variants that outperform the official Jev baseline on multiple benchmarks. Two foundation models were also open-sourced: HyCell, for single-cell representation, and HyPDE, for physical field modeling.
> 
> # How It Performs on Public Leaderboards
> 
> Tencent ran its own set of comparisons. On non-internal public evaluations, Hy4 mostly ranks second to fourth; across twelve main benchmark tables, it does not take first place overall, with one of its best results being second place on E-Bench 2.0. Among open-source models, it is roughly tied with Kimi K3.1 Pro, each winning some: Hy4 leads in coding, 3D modeling, retrieval, and research, while Kimi leads in terminal automation and scientific terminal tasks.
> 
> On the APEX-Agents v1.1 agent benchmark, Tencent's table lists Hy4 at 75.2, slightly above Claude Opus 5.5 (about 74.8) and DeepSeek (about 72.9) as measured by Tencent. However, Opus still leads in most other rows. Tencent itself wrote that there is still a clear gap with the top closed-source models, especially in frontier mathematics, very long development tasks, model iteration, and visual question answering. For example, Hy-FrontierMath T5 is only about 13, far from the near-perfect scores of the GPT series.
> 
> # Pricing, Access, and Safety
> 
> The list pricing is the same as the preview: 6 RMB per million input tokens, 18 RMB per million output tokens, and 0.3 RMB for cache hits. Tencent Cloud TokenHub offers a 50% discount through 2026-12-31. WorkBuddy / CodeBuddy has some free quota from November 4 to 15. It is already available on TokenHub, OpenRouter, WorkBuddy/CodeBuddy, Yuanbao, WeCom's "Dayuan," ima, and more; the weights are "planned to be open-sourced," and a system card will accompany them, but the official version's weights have not been released yet. A cheaper Hy4 flash is also said to be coming soon.
> 
> On safety, the cybersecurity training retains only defense-oriented content, excluding exploitation, reverse engineering, and penetration. On its biosecurity self-test, Hy-BioExperimentBench, it scored about 30.4/100, and the team states it found no significant lowering of barriers to carrying out biological harm.