11-20 更新 实时
首页/ 技术栈/ 正文

Tencent Hy4 上线 Together AI

技术栈 来源:Together AI Blog 2026-11-19 08:00 0 阅读 机器翻译
Tencent Hy4 上线 Together AI

推出 Polymorphic Inference,首先从 Together Compiler 开始 →

推出 Together Post-Training:更好的模型、更快的运行、更低的成本 →

推出 Together Link:在你已使用的运行框架内运行开放模型 →

开源 AI 技术栈:模型、推理、路由器、运行框架与工具 →

GLM-5.3 与 GLM 5.3 Flash 在 DeepSWE 上的对比 →

⚡ Together GPU 集群现已提供按需 B200 →

以 API 形式提供的高性能推理

面向批处理工作负载的推理

基于 Token 的容量,附带 SLA

专用模型推理

在定制硬件上推理

专用容器推理

面向自定义模型的推理

在您的 harness 中运行开放模型

随时间不断优化的推理

探索顶尖开源模型

大规模可靠的 GPU 集群

前沿规模的自定义基础设施

为 AI 构建开发环境

安全存储模型权重与数据

定义你自己的 RL 和 SFT 循环

启动托管的 SFT 和 DPO 作业

微调顶级开源模型

面向生产级 AI 的系统研究

我们的全部研究出版物

Together AI 的技术文档

我们的最新新闻与博客文章

我们的开源演示应用

实用实施指南

用开放模型构建更优方案

在你的工具链中运行开放模型

浏览我们的活动日历

开源 ROI 计算器

以 API 形式提供的高性能推理

面向批处理负载的推理

基于 Token 的容量,附带 SLA

专用模型推理

在定制硬件上运行推理

专用容器推理

面向自定义模型的推理

在你的运行框架中运行开放模型

随时间不断变好的推理

探索顶尖开源模型

规模化部署的可靠 GPU 集群

前沿规模的自定义基础设施

为 AI 构建开发环境

安全存储模型权重与数据

定义你自己的 RL 与 SFT 训练循环

启动托管的 SFT 和 DPO 任务

微调顶级开源模型

面向生产级 AI 的系统研究

我们的全部研究论文

Together AI 技术文档

我们的最新资讯与博客文章

我们的开源演示应用

实用实施指南

用开放模型构建更出色的成果

在你的 harness 中运行开放模型

浏览我们的活动日历

开源 ROI 计算器

Tencent Hy4 已在 Together AI 上线

Tencent 的旗舰模型如今不仅能理解文本,也能理解图像,开发者可以在 Together Serverless Inference 上运行它。

模型:tencent/Hy4 运行于无服务器推理,具备 1M Token 上下文窗口,支持文本和图像输入。

价格:每 1M 输入 Token 0.83 美元,每 1M 输出 Token 2.50 美元,每 1M 缓存输入 Token 0.042 美元。

能力:图像输入、工具调用和 JSON 输出。

从今天起,开发者可以在 Together Serverless Inference 上运行腾讯 Hy4。Hy4 是腾讯的旗舰模型,也是 Hy 系列中首个支持在文本之外接受图像输入的模型。

我们给它布置了两个小任务,看看它在实践中表现如何:从我们的一张图表中提取数据并用 HTML 重建该图表,以及根据单条提示词编写一个浏览器游戏。以下是它返回的结果,以及你可以自己发起图表请求的代码。

它读取了我们的图表,然后重建了它

我们从视觉能力入手,因为这是 Hy4 新增的能力。我们取来 Together Storage Accelerator 博文中的这张图表,把它的截图交给 Hy4,并要求它以 JSON 格式输出其中的内容。

返回结果给出了标题和全部三个数值,与图表中显示的完全一致:

{ "title": "A warm TSA read is 3.2 GB/s: 20x the R2 origin", "bars": [ { "label": "TSA warm read", "value": 3.18 }, { "label": "R2 origin, direct", "value": 0.16 }, { "label": "TSA cold read", "value": 0.14 } ] }

接下来,我们要求它用同一张截图把该图表复现为一个单独的 HTML 文件。这一次,输出既要保留数值,也要保留视觉布局。上方是原图,下方是 Hy4 的版本。

Hy4 复现了布局、标签、数值、配色、虚线顶部界线以及脚注。我们要求它使用通用字体,所以字形有所不同。

在第二项测试中,我们给 Hy4 的不是图片,而是一段文本提示,要求它用一个 HTML 文件写出一个浏览器游戏,不使用任何库或外部资源:

用单个 HTML 文件做一个小巧、原创的浏览器游戏,不使用任何库,也不使用任何外部资源。要有键盘操控、分数、失败条件和重新开始的功能。它应该能在两分钟内带来乐趣。只回复 HTML。

Hy4 返回了一个名为 Polarity 的游戏。我们把回复保存为 HTML 文件,未改动代码,直接在浏览器中打开。

方块从屏幕顶部落下。你要接住与自己颜色相同的方块,并按空格键切换颜色。黄色危险物一旦碰到你就会结束本局。Hy4 还加入了一个讲解规则的标题界面,以及可跨回合延续的最高分,这两项我们都没有要求。

我们完整玩了一局,到达游戏结束界面,并成功重新开始。

Hy4 是 Hy4 Preview 的后续版本,新增了图像输入和 computer-use 能力。腾讯表示,他们训练该模型操作 Blender、CAD 等专业软件。它还与其自有产品团队合作,围绕办公工作、金融研究以及在 Unreal Engine、Unity 和 Godot 中的游戏开发,对 Hy4 进行训练和评估。腾讯的公告中有详细说明。

Hy4 现已可用,模型标识为 tencent/Hy4,通过 Together 的 chat completions API 调用。下面的示例就是本文前面提到的图表请求。要运行它,请安装 Together Python SDK(pip install together),设置 TOGETHER_API_KEY,并把我们使用的图表保存为 chart.png。

import base64 from together import Together client = Together() with open("chart.png", "rb") as f: image = base64.b64encode(f.read()).decode() response = client.chat.completions.create( model="tencent/Hy4", messages=[ { "role": "user", "content": [ { "type": "text", "text": "Read this chart. Return a JSON object with a title " "and a list of bars, each with a label and a value.", }, { "type": "image_url", "image_url": {"url": f"data:image/png;base64,{image}"}, }, ], } ], response_format={"type": "json_object"}, ) print(response.choices[0].message.content)

如果你想在写任何代码之前先试用 Hy4,可以在 playground 中打开它,模型页面则有完整详情。

在 Together AI 上运行 Tencent Hy4

100 万 Token 上下文窗口,支持文本与图像输入、工具调用和 JSON 输出,基于无服务器推理。

Tencent Hy4 在 Together AI 上收费多少?

Tencent Hy4 在 Together AI 上的价格为每 100 万输入 Token 0.83 美元、每 100 万输出 Token 2.50 美元、每 100 万缓存输入 Token 0.042 美元。

Tencent Hy4 在 Together AI 上的上下文窗口是多少?

Tencent Hy4 在 Together AI 上拥有 100 万 Token 的上下文窗口,使用模型字符串 tencent/Hy4 进行调用。

Tencent Hy4 接受图像输入吗?

是的,Tencent Hy4 在 Together AI 上接受文本和图像输入,并返回文本,它是腾讯 Hy 系列中首个具备视觉能力的模型。你需要把每张图片作为 image_url 块放入消息内容中,如上方示例所示。

Tencent Hy4 支持 JSON 输出和工具调用吗?

是的,Tencent Hy4 在 Together AI 上两者都支持。将 response_format 设置为 {"type": "json_object"} 即可获得 JSON 输出;在标准的 tools 参数中传入你的函数即可启用工具调用。

Tencent Hy4 已在 Together AI 上线

Meta Muse Spark Open 已在 Together AI 上线

GLM-5.3 与 GLM-5.3 Flash 在 DeepSWE 上的对比:成本、编程与路由

GLM-5.3 与 GPT-5.6 Sol 在 DeepSWE 上的对比:成本、编程与路由

GLM-5.3 对比 Claude Fable 5:DeepSWE 上的成本、编码与路由

DeepSeek V4 Pro 0813 对比 GPT-5.6 Sol:DeepSWE 上的成本、编码与路由

DeepSeek V4 Pro 0813 对比 Claude Fable 5:DeepSWE 上的成本、编码与路由

DeepSeek-V4 Flash 0731 对比 GPT-5.6 Luna:DeepSWE 上的成本与编码

Kimi K3:开发者完整指南

Kimi K3 对比 GPT-5.6 Sol:DeepSWE 上的成本、编码与路由

Kimi K3 对比 Claude Fable 5:DeepSWE 上的成本与编码

推出 voice finder —— 一款新工具,可从 600+ 种语音中快速为你的应用找到合适的语音

DeepSeek-V4 Pro 现已在 Together AI 上线

Together AI 在 Day 0 为开发者带来 NVIDIA Nemotron 3 Nano Omni

Wan 2.7 视频模型套件现已在 Together AI 上线

Deepgram 语音转文本与语音模型现已在 Together AI 上原生提供

在 Together AI 上构建实时语音智能体

Together AI 在 Day 0 为开发者带来 NVIDIA Nemotron 3

Rime Arcana V3 Turbo 和 Rime Arcana V3 现已在 Together AI 上线

如何为生产环境选择合适的开放模型

MiniMax Speech 2.6 Turbo 现已在 Together AI 上原生可用

Rime 语音模型现已在 Together AI 上可用

宣布 NVIDIA Nemotron 3 Nano 原生可用,这是 NVIDIA 最新的推理模型

FLUX.2:多参考图像生成现已在 Together AI 上可用

宣布面向实时语音 AI 智能体的最快推理

如何评估和基准测试大语言模型(LLM)

将 Together AI 模型库扩展至多媒体生成领域,新增 40+ 图像和视频模型

DeepSeek-V3.1:混合思考模型现已在 Together AI 上可用

OpenAI 全新开放模型 gpt-oss 与 o4-mini 对比:一次真实场景实测

OpenAI 的开放模型现已在 Together AI 上线

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Tencent Hy4 is live on Together AI)

Introducing Polymorphic Inference, starting with Together Compiler →

Introducing Together Post-Training: Better models, faster runs, lower costs →

Introducing Together Link: Run open models inside the harness you already use →

The Open Source AI Stack: models, inference, routers, harnesses & tools →

GLM-5.3 vs. GLM 5.3 Flash on DeepSWE →

⚡ On-demand B200s now available on Together GPU Clusters →

High-performance inference as APIs

Inference for batch workloads

Token-based capacity with SLAs

Dedicated Model Inference

Inference on custom hardware

Dedicated Container Inference

Inference for custom models

Run open models in your harness

Inference that gets better over time

Explore Top Open-Source Models

Reliable GPU clusters at scale

Custom infrastructure at frontier scale

Build development environments for AI

Store model weights & data securely

Define your own RL and SFT loop

Launch managed SFT and DPO jobs

Fine-tune top open-source models

Systems research for production AI

All our research publications

Technical docs for Together AI

Our latest news & blog posts

Our open-source demo apps

Practical implementation guides

Build better with open models

Run open models in your harness

Explore our events calendar

Open-source ROI calculator

High-performance inference as APIs

Inference for batch workloads

Token-based capacity with SLAs

Dedicated Model Inference

Inference on custom hardware

Dedicated Container Inference

Inference for custom models

Run open models in your harness

Inference that gets better over time

Explore Top Open-Source Models

Reliable GPU clusters at scale

Custom infrastructure at frontier scale

Build development environments for AI

Store model weights & data securely

Define your own RL and SFT loop

Launch managed SFT and DPO jobs

Fine-tune top open-source models

Systems research for production AI

All our research publications

Technical docs for Together AI

Our latest news & blog posts

Our open-source demo apps

Practical implementation guides

Build better with open models

Run open models in your harness

Explore our events calendar

Open-source ROI calculator

Tencent Hy4 is live on Together AI

Tencent's flagship model now understands images as well as text, and developers can run it on Together Serverless Inference.

The model:tencent/Hy4 on serverless inference, with a 1M-token context window and text and image input.

The price: \$0.83 per 1M input tokens, \$2.50 per 1M output tokens, \$0.042 per 1M cached input tokens.

Capabilities: image input, tool calling and JSON output.

Starting today, developers can run Tencent Hy4 on Together Serverless Inference. Hy4 is Tencent's flagship model and the first in the Hy family to accept images alongside text.

We gave it two small jobs to see what it does in practice: extracting the data from one of our charts and rebuilding that chart in HTML, then writing a browser game from a single prompt. Here's what came back, along with the code to make the chart request yourself.

It read our chart, then rebuilt it

We started with vision, since that's the capability Hy4 adds. We took this chart from our Together Storage Accelerator post, gave Hy4 a screenshot of it, and asked for its contents as JSON.

The response returned the title and all three values exactly as they appear in the chart:

{ "title": "A warm TSA read is 3.2 GB/s: 20x the R2 origin", "bars": [ { "label": "TSA warm read", "value": 3.18 }, { "label": "R2 origin, direct", "value": 0.16 }, { "label": "TSA cold read", "value": 0.14 } ] }

Next, we asked it to reproduce the chart as a single HTML file, using the same screenshot. This time, the output needed to preserve the visual layout as well as the numbers. The original is on top, and Hy4's version is below it.

Hy4 reproduced the layout, the labels, the values, the colors, the dashed ceiling line and the footnote. We asked it to use generic fonts, so the typeface differs.

For the second test, we gave Hy4 a text prompt rather than an image and asked it to write a browser game in one HTML file, with no libraries or external assets:

Make a small, original browser game in a single HTML file, with no libraries and no external assets. Keyboard controls, a score, a way to lose and a restart. It should be fun for two minutes. Reply with only the HTML.

Hy4 returned a game called Polarity. We saved the response as an HTML file and opened it in a browser without changing the code.

Blocks fall from the top of the screen. You catch the ones that match your color and press Space to switch colors. A yellow hazard ends the run if it reaches you. Hy4 also added a title screen that explains the rules and a best score that carries over from one round to the next, neither of which we asked for.

We played through a round, reached the game-over screen and restarted successfully.

Hy4 follows Hy4 Preview, adding image input and computer-use capabilities. Tencent says it trained the model to operate professional software such as Blender and CAD tools. It also worked with its own product teams to train and evaluate Hy4 on office work, financial research and game development in Unreal Engine, Unity and Godot. Tencent's announcement has the details.

Hy4 is available now under the model string tencent/Hy4, through Together's chat completions API. The example below is the chart request from earlier in this post. To run it, install the Together Python SDK (pip install together), set a TOGETHER_API_KEY, and save the chart we used as chart.png.

import base64 from together import Together client = Together() with open("chart.png", "rb") as f: image = base64.b64encode(f.read()).decode() response = client.chat.completions.create( model="tencent/Hy4", messages=[ { "role": "user", "content": [ { "type": "text", "text": "Read this chart. Return a JSON object with a title " "and a list of bars, each with a label and a value.", }, { "type": "image_url", "image_url": {"url": f"data:image/png;base64,{image}"}, }, ], } ], response_format={"type": "json_object"}, ) print(response.choices[0].message.content)

If you'd like to try Hy4 before writing any code, you can open it in the playground, and the model page has the full details.

Run Tencent Hy4 on Together AI

1M-token context, text and image input, tool calling and JSON output, on serverless inference.

How much does Tencent Hy4 cost on Together AI?

Tencent Hy4 on Together AI costs \$0.83 per 1M input tokens, \$2.50 per 1M output tokens and \$0.042 per 1M cached input tokens.

What is the context window of Tencent Hy4 on Together AI?

Tencent Hy4 has a 1M-token context window on Together AI, and you call it with the model string tencent/Hy4.

Does Tencent Hy4 accept images?

Yes, Tencent Hy4 accepts text and image input on Together AI and returns text, and it's the first model in Tencent's Hy family with vision. You send each image as an image_url block in the message content, as in the example above.

Does Tencent Hy4 support JSON output and tool calling?

Yes, Tencent Hy4 supports both on Together AI. You get JSON output by setting response_format to {"type": "json_object"}, and you enable tool calling by passing your functions in the standard tools parameter.

Tencent Hy4 is live on Together AI

Meta Muse Spark Open is live on Together AI

GLM-5.3 vs. GLM-5.3 Flash on DeepSWE: Cost, Coding, and Routing

GLM-5.3 vs. GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

GLM-5.3 vs. Claude Fable 5 on DeepSWE: Cost, Coding, and Routing

DeepSeek V4 Pro 0813 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

DeepSeek V4 Pro 0813 vs Claude Fable 5 on DeepSWE: Cost, Coding, and Routing

DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding

Kimi K3: the complete developer guide

Kimi K3 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

Kimi K3 vs Claude Fable 5 on DeepSWE: Cost and Coding

Introducing voice finder — a new tool to quickly find the right voice for your app from over 600+ voices

DeepSeek-V4 Pro now available on Together AI

Together AI Brings NVIDIA Nemotron 3 Nano Omni to Developers on Day 0

Wan 2.7 video model suite now available on Together AI

Deepgram speech-to-text and voice models now available natively on Together AI

Build real-time voice agents on Together AI

Together AI Brings NVIDIA Nemotron 3 to Developers on Day 0

Rime Arcana V3 Turbo and Rime Arcana V3 now available on Together AI

How to choose the right open model for production

MiniMax Speech 2.6 Turbo now available natively on Together AI

Rime voice models now available on Together AI

Announcing native availability of NVIDIA Nemotron 3 Nano, NVIDIA’s latest reasoning model

FLUX.2: Multi-reference image generation now available on Together AI

Announcing the fastest inference for realtime voice AI agents

How to evaluate and benchmark Large Language Models (LLMs)

Expanding Together AI Model Library into multimedia generation with 40+ new image and video models

DeepSeek-V3.1: Hybrid Thinking Model Now Available on Together AI

OpenAI's New Open gpt-oss Models vs o4-mini: A Real-World Comparison

Announcing the Availability of OpenAI's Open Models on Together AI

VirtueGuard: Enterprise-Grade AI Security and Safety Now on Together AI

Qwen3-Coder: The Most Capable Agentic Coding Model Now Available on Together AI

Kimi K2: Leading Open-Source Model Now Available on Together AI

Together AI Launches Speech-to-Text: High-Performance Whisper APIs

FLUX.1 Kontext models: Character consistency and precise image editing without fine-tuning

Together AI partners with Meta to offer Llama 4: SOTA Multimodal MoE Models

How to deploy DeepSeek-R1 and distilled models securely on Together AI

Mistral Small 3 API now available on Together AI: A new category leader in small models

Build ultra low latency voice AI applications with Together AI and Cartesia Sonic

Announcing Llama 3.3 70B, with enhanced reasoning, mathematics, and instruction-following on Together AI

FLUX Tools now available via Together APIs: Get greater control over image generation with Canny, Depth and Redux models

How to build a real-time image generator with Flux and Together AI

FLUX API is now available on Together AI: New FLUX1.1 [pro] and free access to FLUX.1 [schnell]

Together AI launches Llama 3.2 APIs for vision, lightweight models & Llama Stack: powering rapid development of multimodal agentic apps

Introducing Together Rerank API and exclusive access to Salesforce LlamaRank model for enhanced enterprise search

Llama 3.1: Same model, different results. The impact of a percentage point.

Together AI partners with Meta to release Llama 3.1 models for inference and fine-tuning with accelerated performance at full accuracy

Together AI partners with Meta to release Meta Llama 3 for inference and fine-tuning

Can you feel the MoE? Mixtral available with over 100 tokens per second through Together Platform!

Releasing GPT-JT powered by open-source AI

Tencent Hy4 is live on Together AI

Introducing Together Storage Accelerator: fast, free data caching for your GPU cluster

Last year we built megakernels. This year we deleted them: The road to Together Compiler

Making open models work for long-horizon agentic workloads

How we do agent evals at Together AI

Together Compiler: Continual Compilation for AI inference

From Static to Polymorphic Inference: A new era for production open-source inference systems

Auto repair for Together GPU Clusters: from GPU fault to healthy node

Together Post-Training: frontier-grade model shaping, the way that fits you

Meta Muse Spark Open is live on Together AI

Expanding our enterprise inference capacity with IBM Cloud and NVIDIA

Together Link: frontier-quality open models in the harness you already use

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 Together AI Blog(Tencent Hy4 is live on Together AI),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗