What do the sources establish?
| Platform / Source | Core Stance / Quote | Engagement Metric | Timestamp |
|---|---|---|---|
| Tencent releases its Hy4 open-weight foundation model (Techmeme, citing Bloomberg) | Reported context: [object Object] [object Object] | Not recorded | |
| Tencent launches Hy4 with vision and computer-use capabilities | Reported context: [object Object] [object Object] [object Object] [object Object] | Not recorded |
THE SHORT VERSION
- Tencent released Hy4 on November 4, a 770-billion-parameter foundation model it says joins the top tier of Chinese open-weight AI systems for the first time.腾讯于 11 月 4 日发布 Hy4，这是一个 7700 亿参数的基础模型，公司称其首次进入中文开源权重模型的第一梯队。
- The model adds visual understanding and computer-use abilities, keeps preview API pricing of $0.834 per million input tokens, and is being rolled into Tencent's WorkBuddy and CodeBuddy products.该模型新增视觉理解与电脑操作能力，API 价格维持在每百万输入 token 0.834 美元，并已接入腾讯的 WorkBuddy 与 CodeBuddy。
- Tencent says it will publish the weights and a system card shortly, so the open-weight claim is currently a stated plan rather than a downloadable model.腾讯表示将很快公布权重与系统卡，因此“开源权重”目前仍是一项承诺，而不是可下载的模型。
A model pitched at the top tier一款自称第一梯队的模型
Tencent released Hy4 on Wednesday, a foundation model the company says joins the top tier of Chinese open-weight AI systems for the first time. Techmeme summarised the release from Bloomberg, which also reported that a smaller Flash version is planned. The top-tier claim is Tencent's own, and the open weights that would let outsiders check it are still to come.腾讯周三发布 Hy4，称这是首个进入中文开源权重模型第一梯队的基础模型。Techmeme 汇总了彭博社的报道，并提到腾讯还计划推出更小的 Flash 版本。第一梯队的说法来自腾讯自己，而让外界得以检验它的权重，目前还没有放出来，这一点值得记住。
The framing matters more than the adjective. Chinese labs have spent two years arguing that open weights can match closed models on real work, and each release now arrives with a comparison table attached. Tencent is asking buyers to treat Hy4 as a peer of the models it is racing, not as the cheaper substitute in the category.这层措辞比形容词本身更重要。过去两年，中国实验室一直在论证开源权重能在真实工作中追上闭源模型，于是每次发布都附带一张对比表。腾讯希望买家把 Hy4 当作它正在追赶的那些模型的同级对手，而不是这一档里的廉价替代品。
What Hy4 actually doesHy4 实际能做什么
TechNode reported that Hy4 adds visual understanding and computer-use capabilities to Tencent's flagship AI model. It can interpret charts and documents, reason across multiple images, and operate command-line tools and graphical interfaces, the outlet reported. Training covered applications such as Blender and CAD software, alongside tasks that span several programs at once. Tencent is selling this as a tool for desk work rather than a chat demo.TechNode 报道称，Hy4 为腾讯的旗舰模型加入了视觉理解与电脑操作能力。据该媒体报道，它可以解读图表与文档、在多张图片之间推理，并操作命令行工具与图形界面。训练覆盖了 Blender 与 CAD 等应用，也覆盖需要同时操作多个程序的任务，取向明显偏向办公场景，而不是聊天演示。
The 770-billion-parameter size puts Hy4 in the large-model class, and TechNode reported that it is available across Tencent products and through third-party platforms including OpenRouter. Tencent said development focused on office tasks, financial analysis and game development, with product teams helping train and evaluate the model against real workflows rather than synthetic benchmarks, an arrangement vendors rarely document in this detail.7700 亿参数让 Hy4 进入大模型行列。TechNode 报道称，它已在腾讯各产品中上线，也可通过 OpenRouter 等第三方平台调用。腾讯表示研发重点在办公任务、财务分析与游戏开发，产品团队参与训练与评估，用真实工作流而非合成基准来检验模型，这种参与方式厂商很少写得这么细。
The price held, the claims moved价格没动，说法升级
TechNode reported that standard API pricing remains unchanged from the Hy4 preview at $0.834 per million input tokens and $2.501 per million output tokens, with a 50% discount on Tencent Cloud TokenHub through December 31. Holding the price while raising the capability claim is the competitive move here, not the benchmark table. A discount that runs to the end of the year is also a deadline for buyers to test the model.TechNode 报道，标准 API 价格与 Hy4 预览期持平：每百万输入 token 0.834 美元、每百万输出 token 2.501 美元，并在 12 月 31 日前于腾讯云 TokenHub 提供五折优惠。在提高能力主张的同时按住价格，才是这里真正的竞争动作，而不是那张基准表。折扣持续到年底，也给买家设下了一个试用期限。
Cheap tokens are how a model buys adoption before its weights ship. This site has tracked the same pressure from the supply side in the inference economics behind fal's valuation, where the argument is that inference margins, rather than model quality alone, decide which vendors hold their ground. A model that is cheap to call is easier to design around than one that tops a table.在权重发布之前，便宜 token 是模型换取采用率的方式。本站从供给端追踪过同样的压力，见fal 估值背后的推理经济学：决定谁能站住的，是推理毛利，而不只是模型质量。调用便宜的模型，比排名第一的模型更容易被设计进产品里。
Weights, and the promise of them权重，以及关于权重的承诺
TechNode reported that Tencent plans to release its weights and a system card shortly, which is the step that turns a hosted product into an open-weight model that other people can run. Until that happens, the open part of the announcement rests on a stated intention rather than a downloadable file. Buyers who need a model inside their own network cannot plan around an intention.TechNode 报道，腾讯计划很快公布权重与系统卡，这一步才把托管产品变成别人可以自己运行的开源权重模型。在那之前，发布中“开源”的部分仍只是意向，而不是一个可下载的文件。需要在自有网络里跑模型的买家，没法围绕一个意向做规划。
The sequence is now standard. Labs ship an API first, watch how the model behaves under real traffic, then publish weights once the safety write-up is finished. It buys time, it keeps the announcement cycle moving, and it lets the company correct a model before thousands of people fine-tune a copy of it, which is a reasonable order of operations for a model this large.这个顺序如今已是惯例。实验室先上线 API，观察模型在真实流量下的表现，等安全评估写完再放出权重。这样能争取时间，让发布节奏不断档，也能在成千上万人微调副本之前先修正模型，对一个这个体量的模型来说，这个次序是合理的。
Where the open-weight race stands开源权重竞赛的位置
Hy4 lands in a Chinese open-weight field that already includes DeepSeek, Z.ai and Alibaba, and comparison has become part of the product. Each release is measured against the others within days, on agentic tasks, coding and cost per token, by users who publish their own tests and are quick to say when a model disappoints. Nobody gets a quiet launch any more.Hy4 进入的是一个已有 DeepSeek、Z.ai 与阿里巴巴的中文开源权重赛场，而对比本身已成为产品的一部分。每次发布都会在几天内被拿来与对手比较：智能体任务、编程、每 token 成本，测试由用户自己跑，也会直言模型的失望之处。已经没有人能安静地发布模型了。
For buyers, that competition has been useful. Credible open weights set a ceiling on hosted inference prices and give companies a fallback when a vendor changes terms, which is why procurement teams now ask about licence conditions before they ask about benchmark scores. A model that can be self-hosted is a negotiating position as much as a technical option.对买家来说，这种竞争是有用的。可信的开源权重为托管推理价格设了上限，也让企业在供应商改条款时有退路，这正是采购团队现在先问许可证条件、再问基准分数的原因。能自己部署的模型，既是技术选项，也是一种谈判筹码。
Why Tencent wants this fight腾讯为何要打这一仗
Distribution is Tencent's real argument. Hy4 went into WorkBuddy and CodeBuddy with a limited free usage quota from November 4 to November 15, TechNode reported, which puts the model in front of office and coding users without a separate sales motion or a developer conference. Existing products are a cheaper route to users than new ones, and the quota sets a deadline for teams to try the model.腾讯真正的论据是分发能力。据 TechNode 报道，Hy4 已接入 WorkBuddy 与 CodeBuddy，并在 11 月 4 日至 15 日提供有限免费额度，这让模型直接出现在办公与编程用户面前，不需要另开一场销售动作或开发者大会。现成产品比新建渠道更省钱，额度也给了团队一个试用期限。
Tencent also named game development as a target workflow, a corner of the market this site covered when Anthropic moved against game cloning. Owning the tools developers already have open is worth more, commercially, than winning a leaderboard for a week. Studios choose engines and assistants for years at a time, not for a single release cycle.腾讯还把游戏开发列为目标工作流，这正是本站报道Anthropic 出手限制游戏克隆时涉及的领域。从商业上看，占住开发者已经打开的工具，比赢下一周的排行榜更有价值。工作室选引擎和助手是按年算的，不是按一次发布算的，这一点常被忽略。
What to watch next接下来看什么
The first test is the weights release, and the second is whether independent evaluations reproduce the top-tier claim. A system card will describe what Tencent tested and how; outside runs will show what it left out, which is usually the part buyers care about most. Neither is available on release day, which is why the claim reads as a promise.第一个检验是权重发布，第二个是独立评测能否复现“第一梯队”的说法。系统卡会说明腾讯测了什么、怎么测；外部复跑则会显示它漏了什么，而那通常是买家最在意的部分。这两样在发布当天都拿不到，所以那句话目前只能当作承诺来读。
The third is price. If Hy4 holds its rate while rivals cut theirs, the TokenHub discount becomes the story rather than the model. This site has followed the gap between AI spending and returns long enough to expect that arithmetic, rather than a benchmark margin, to decide the next round of Chinese releases. A price cut from any of the three rivals would show up in that arithmetic within a quarter.第三个是价格。如果对手降价而 Hy4 守住自己的费率，那么真正值得写的就是 TokenHub 折扣，而不是模型本身。本站长期追踪AI 支出与回报之间的落差，也因此预期，决定下一轮中文发布的会是这笔算术，而不是基准表上的几个点。对手只要降价，一个季度内就能在账单上看出来。
Sources & reporting limits来源与核查范围
Follow Models to find related stories on your next visit.关注模型动态，下次访问时阅读相关文章。