Video Models Finally Get Their Own Codex: MiniMax Design Says Goodbye to Frame-by-Frame Editing
Bilingual Edition · English First | 中文版见下方
If you’re reading this for the English version, keep scrolling after the English section for the complete Chinese edition.
On August 20, MiniMax officially launched MiniMax Design, a multimodal creative agent workspace designed to turn video creation into an intelligent, end-to-end workflow.
The timing is notable. The release came less than three weeks after the open-source release of H3.
In just a few weeks, MiniMax moved from an open-source video model to a commercial production workflow. This is more than ordinary productization. MiniMax Design is being positioned as “the Codex for video models.”
And that comparison is actually quite revealing.
Codex changed software development by moving developers away from manually manipulating every line of code toward expressing what they wanted to accomplish.
Design is attempting something similar for video: moving creators away from manipulating every frame toward manipulating creative intent.
From Manipulating Pixels to Manipulating Intent
To understand what Design changes, it helps to look at the traditional AI-video workflow.
It is fragmented.
A creator might write a prompt in one AI assistant, copy it into a video-generation tool, generate a clip, discover that something is wrong, return to the prompt, modify it, regenerate the video, export the result into another editing application, and repeat the process.
Prompt writing, frame editing, asset management, version control, and tool switching all become part of the workflow.
The model may be intelligent, but the production process remains complicated.
MiniMax Design takes a different approach.
Instead of requiring users to specify every individual operation, the user describes the desired outcome in natural language. The agent then interprets the goal, decomposes the task, identifies the necessary assets, selects appropriate models and skills, and coordinates the workflow from generation to editing and delivery.
The key interface is therefore no longer a traditional timeline or a collection of technical controls.
It is language.
Creators do not necessarily need to explain exactly how every pixel should change. They can describe the scene, characters, movement, style, or desired result, while the system determines how those requirements should be translated into executable operations.
MiniMax describes this shift as moving from “manipulating pixels” to “manipulating semantics and intent.”
That distinction could become one of the most important changes in AI-assisted video production.
H3 Is the Engine. The Agent Is the Brain.
Underneath Design sits MiniMax’s native multimodal video model, H3.
H3 is designed to understand multimodal context involving text, images, video, and audio within a unified generation framework. It can generate video at resolutions of up to 2K, with clips of up to 15 seconds and native stereo audio.
But the model itself is only one part of the system.
The more interesting layer is the agent.
MiniMax Design functions as a kind of harness around the underlying model. Instead of forcing users to understand the model’s capabilities, input formats, parameter requirements, and limitations, those operational details can be embedded into agents and skills.
When a creator asks for a particular video, Design can determine:
What text information is required
Which images or reference assets are needed
Whether existing video or audio should be preserved
Which generation capabilities should be invoked
How the resulting materials should be assembled
What needs to be modified when the first result is not satisfactory
This is similar to what happened when advanced reasoning models were paired with coding agents such as Codex.
The model itself provides intelligence.
The agent provides workflow intelligence.
MiniMax Design is attempting to turn H3 from a video-generation model into something much closer to a production system.
The 3D Director: Moving Trial and Error Before the Storyboard
One of Design’s more interesting features is its 3D Director.
Instead of immediately generating the final video, users can describe the desired scene and camera movement in natural language. Design can then preview the composition and spatial relationships inside a 3D environment.
Only after the scene is confirmed does the system generate the final video.
That changes where the trial-and-error process happens.
In traditional video production, discovering that a character is standing in the wrong position or that the camera movement does not work can mean reshooting or rebuilding an entire sequence.
With the 3D Director, much of that experimentation can happen before final generation.
Consider a simple scenario:
A female protagonist and a male protagonist are walking across a grassland. They encounter a large river. She jumps across the river while he continues walking along the bank.
The creator does not necessarily need to manually construct every shot.
Design can first generate a director-stage preview, allowing the creator to inspect character positioning, camera placement, movement, and spatial relationships.
If something looks wrong, the creator can adjust the characters or camera and regenerate the scene.
Only after the composition works does the system move toward the final video.
The value here is not simply “text-to-video.”
It is the conversion of ambiguous creative language into visible spatial references that humans can still intervene in.
That makes the process more controllable, iterative, and production-oriented.
From Generation Tool to Commercial Workflow
Design is currently aimed at use cases including e-commerce, education, and creative short-form video.
For brands, the attraction is obvious: large volumes of marketing assets can potentially be generated without rebuilding every video manually.
For educators, lesson plans and instructional material can be transformed into short-form visual content.
For creators, the bigger opportunity may be workflow consolidation.
Instead of switching between multiple specialized tools, the creator can increasingly treat the agent as the coordination layer.
Design also integrates with existing open-source workflows such as ComfyUI, allowing creators to retain greater flexibility over parameters and production pipelines.
Local ComfyUI workflows can potentially be incorporated into the broader Design environment, allowing the agent to assist with nodes, parameters, and workflow adjustments.
That creates an interesting ecosystem strategy:
The open-source community experiments. The agent turns those experiments into reusable production capabilities.
The competitive advantage therefore does not necessarily come only from the underlying video model.
It can also come from the workflow ecosystem built around it.
English Version — Part 2
The Hidden Variable Behind Video AI: Compute Cost
There is, however, another side to the equation.
No matter how smooth the creative workflow becomes, every generated frame ultimately consumes computing resources.
H3’s video-generation pricing has been cited at approximately RMB 0.8 per second for 2K video.
That number may look manageable for an individual experiment.
But commercial production operates on a completely different scale.
A single creator might generate dozens of clips while testing an idea. A marketing team may generate hundreds of variations. A large enterprise could run thousands of generation tasks every day.
As AI video moves from experimentation into production, compute consumption can grow dramatically.
That is where the infrastructure layer becomes increasingly important.
Where Unexhub Fits Into the Picture
Unexhub is positioned around a different part of the AI production stack: token aggregation, distribution, and compute trading.
Its goal is to move AI compute procurement away from a simple fixed-price model and toward a more dynamic marketplace structure.
Three mechanisms are particularly important.
1. Real-Time Settlement
For smaller AI suppliers, cash flow can be almost as important as pricing.
Under conventional settlement cycles, a supplier may provide API capacity today but wait much longer to receive payment.
Unexhub’s model aims to provide real-time settlement, allowing suppliers to recover revenue as usage occurs.
That can reduce the working-capital pressure associated with selling AI inference capacity.
2. Early-Access Zero Commission
For suppliers entering the ecosystem during the early stage, zero commission can reduce the initial cost of participation.
Instead of paying a platform fee while trying to build demand, early suppliers can focus on connecting capacity, attracting users, and validating their commercial model.
For a marketplace, this also has a strategic purpose:
Lowering the supply-side barrier makes it easier to build liquidity.
3. API Bidding
The most interesting mechanism is dynamic API bidding.
Instead of treating model pricing as a permanently fixed number, prices can respond to supply and demand.
When capacity is abundant, users may have access to more competitive prices.
When demand increases, suppliers have an opportunity to capture higher returns.
This creates a market mechanism rather than simply another API price list.
For developers, the potential benefit is straightforward: pay according to actual token consumption while searching for more competitive inference prices.
Unexhub states that developers can access participating AI models and, under applicable offers, potentially obtain prices as low as 30% of prevailing market prices.
For suppliers, the corresponding value is on the other side of the transaction: one-click channel integration and real-time settlement of usage revenue.
The two sides reinforce each other.
Lower procurement costs attract developers.
More developer demand attracts suppliers.
More suppliers improve competition.
More competition creates better price discovery.
That is how an AI compute marketplace can potentially develop liquidity.
The Cost Structure of AI Video Is Being Rewritten
MiniMax Design and Unexhub operate at different layers, but they point toward the same structural change.
At the application layer, Design reduces the human cost of video production.
At the infrastructure layer, marketplace-based compute distribution attempts to reduce the cost of AI inference.
The result is a two-sided optimization:
Humans spend less time operating tools. Machines spend less money performing computation.
That distinction matters.
For years, AI progress was mostly discussed in terms of model intelligence.
Then the industry started talking about token prices.
Now another question is becoming increasingly important:
How much does it cost to complete an actual piece of work?
That is a much more useful metric for commercial AI.
A video creator does not fundamentally care whether a model consumes 1 million or 10 million tokens.
They care whether the system can produce a usable commercial video for a predictable cost.
A software engineer does not fundamentally care how many tokens an agent consumed.
They care whether the agent successfully completed the task.
The industry is therefore gradually moving from token economics toward task economics.
Video Creation Is Moving Up the Abstraction Stack
The historical video-production workflow was built around specialized roles:
Scriptwriter
Storyboard artist
Director
Cinematographer
Editor
Sound designer
AI does not necessarily eliminate these creative functions.
Instead, it changes how they are accessed.
With Design, a creator can increasingly start with an idea rather than a technical workflow.
You do not necessarily need to know how to construct every ComfyUI node.
You do not necessarily need to understand every H3 parameter.
You do not necessarily need to move between five different applications.
You need to explain what you want.
The system increasingly handles how to make it happen.
That is the same abstraction shift that made coding agents so powerful.
Video Models Finally Have Their Own Codex
The significance of MiniMax Design therefore goes beyond another AI video-generation interface.
It represents a broader transition:
from models to agents,
from prompts to workflows,
from pixels to semantics,
from generation to production.
And once those workflows become commercially viable, compute economics become the next battlefield.
MiniMax Design is trying to make video creation easier.
Unexhub is trying to make the underlying AI compute market more competitive, flexible, and transparent.
The combination points toward a future in which creators no longer think primarily about models, APIs, or infrastructure.
They simply describe the result they want.
The system determines the workflow.
And the marketplace determines the most efficient way to pay for the intelligence required to produce it.
Video models now have their own Codex. The next question is whether semantic-level creation can move from impressive demonstrations into everyday commercial production—and whether a more competitive compute market can make that transition economically sustainable.
中文版 Part 1
视频模型也有了自己的 Codex:MiniMax Design 告别逐帧编辑
8 月 20 日,MiniMax 正式发布多模态创作智能体工作台 MiniMax Design。
距离 H3 开源不到三周。
短短几周时间,从一个开源视频模型,到一个面向商业生产的完整工作流,MiniMax 做的已经不只是传统意义上的“产品化”。
它试图解决的是另一个问题:
如果 Codex 能让程序员从“写代码”转向“描述要完成的任务”,那么视频模型是不是也可以让创作者从“操作画面”转向“表达创意”?
MiniMax Design 给出的答案是:可以。
因此,它被称为“视频模型界的 Codex”。
从“操纵像素”到“操纵语义与意图”
先看传统 AI 视频创作流程。
它往往是碎片化的。
用户可能先在一个 AI 工具里写提示词,再复制到视频生成平台;生成之后发现人物动作不对,又回到原来的对话框修改提示词;生成满意的视频后,还要进入剪辑软件处理素材、调整镜头、管理版本。
模型越来越聪明,但整个生产流程依然很复杂。
真正消耗时间的,往往不是生成本身,而是工具之间的反复切换。
MiniMax Design 的思路完全不同。
用户只需要描述最终想实现什么,智能体负责理解目标、拆解任务、识别所需素材,并调用相应的模型和技能,完成从素材处理、内容生成、剪辑到最终交付的一整套流程。
核心交互界面因此从传统的时间轴和参数面板,逐渐转向了:
自然语言。
创作者不需要告诉系统每一个像素应该移动多少,也不一定需要掌握每一个参数的含义。
只要把人物、场景、动作、镜头、风格和最终效果描述清楚,Design 就可以负责把这些创作意图转化成具体的执行步骤。
MiniMax 对这种变化的概括非常准确:
从“操纵像素”,转向“操纵语义与意图”。
这可能正是 AI 视频进入商业生产之后最重要的变化之一。
H3 是引擎,Agent 才是大脑
Design 的底层基础是 MiniMax 原生多模态视频模型 H3。
H3 可以统一理解文本、图片、视频和音频等多模态上下文,并生成最高 2K 分辨率、最长 15 秒、支持原生立体声的视频内容。
但如果只看 H3,本质上它仍然只是一个强大的视频生成模型。
Design 真正有意思的地方,在于 Agent 层。
MiniMax Design 可以被理解为围绕模型建立的一套 Harness(应用框架)。
模型负责提供能力。
Agent 负责理解任务、组织能力、调用工具和协调流程。
面对一个具体的视频创作需求,Design 可以判断:
需要哪些文本信息;
需要哪些图片或参考素材;
哪些视频和音频内容需要保留;
应该调用哪些模型能力;
如何组合生成结果;
第一版不满意之后应该修改什么。
这和 GPT-5.6 Sol 配合 Codex 的逻辑非常相似。
模型本身提供智能,Agent 把智能变成工作流。
Design 做的,就是把 H3 从一个“视频生成引擎”,进一步变成一个能够参与生产流程的智能系统。
3D 导演台:把试错放到最终生成之前
Design 里面一个很有意思的功能,是 “3D 导演台”。
用户用自然语言描述场景、人物和镜头需求后,系统可以先在 3D 空间里预演构图、人物关系和镜头运动。
确认没有明显问题之后,再让 H3 生成最终视频。
这实际上改变了 AI 视频创作中“试错”的位置。
传统方式往往是:
先生成 → 发现问题 → 修改提示词 → 再生成。
如果问题比较严重,就可能需要重新生成整个镜头。
而 3D 导演台把部分试错成本提前到了故事板阶段。
例如:
“女主和男主在草原上行走,遇到一条大河。她纵身跃过河流,他沿岸大步前行。”
过去,创作者可能需要反复调整提示词,直到人物位置、动作和镜头关系基本正确。
而在 Design 中,可以先通过导演台查看场景预览。
如果人物站位不对,可以调整角色位置。
如果镜头角度不对,可以调整摄像机。
如果动作关系不合理,可以重新规划场景。
确认之后,再生成最终视频。
所以,3D 导演台真正解决的并不是“文生视频”本身。
它解决的是:
如何把模糊的自然语言创意,变成可视化、可干预、可迭代的空间关系。
这意味着 AI 视频开始从“黑盒生成”,走向“可控生产”。
中文版 Part 2
从生成工具到商业生产工作流
目前,MiniMax Design 重点覆盖电商、教育、创意短视频等应用场景。
对于品牌而言,它可以帮助团队快速批量制作营销素材。
对于教育行业,可以将课程内容、教案和知识点进一步转化成短视频。
而对于创作者来说,更重要的可能是:
减少工具切换。
过去,一个完整的视频生产流程可能需要多个专业软件协同。
现在,Agent 正逐渐成为这些工具之间的“协调层”。
Design 同时支持与 ComfyUI 等现有工作流进行集成,让创作者能够保留原有的专业能力和自定义空间。
本地 ComfyUI 工作流也可以进一步接入 Design,由智能体辅助调整节点、参数和流程。
这实际上形成了一种很有意思的生态逻辑:
开源社区负责探索,Agent 负责把探索沉淀成可复用的生产能力。
因此,Design 的竞争力不一定只来自 H3 本身。
它还可能来自围绕 H3 建立起来的工作流生态。
底层真正的变量:算力成本
不过,视频创作效率提升之后,还有一个无法绕开的现实:
每一帧视频背后都是算力。
H3 的视频生成价格此前被公布为约 0.8 元/秒(2K 分辨率)。
对于个人测试来说,这个价格并不算夸张。
但商业生产完全是另一回事。
一个创作者可能为了一个创意生成几十个版本。
一个营销团队可能每天生成数百条素材。
大型企业甚至可能同时运行数千个视频生成任务。
当 AI 视频从“尝鲜工具”真正进入商业生产线之后,算力消耗也会快速放大。
于是,AI 视频的下一场竞争就不只是模型能力竞争。
还包括:
谁能以更低、更稳定、更透明的成本获得算力。
这正是 Unexhub 所瞄准的基础设施层。
Unexhub:把算力采购从固定价格推向动态市场
Unexhub 的定位,是 Token 聚合、分发与算力交易中心。
它试图改变传统 AI API 的采购方式:
过去,开发者看到的是厂商固定价格。
未来,则可能通过市场机制寻找更具竞争力的价格。
Unexhub目前聚合不同规模的 AI 供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型供应商,同时也接入个人和中小企业提供的 AI 模型能力。
真正值得关注的是它设计的三个机制。
① 即时结算:让供应商更快回款
传统 API 分销模式中,中小供应商提供算力之后,往往需要等待结算周期。
Unexhub希望通过 实时结算,让供应商根据实际调用情况更快获得对应收入。
对于规模较小的算力供应商而言,这意味着资金周转压力可以明显降低。
算力卖出去,就能够更快形成现金流。
② 早期零佣金:降低供应侧进入门槛
平台早期对入驻供应商提供 零佣金权益,核心价值并不只是“少交一笔平台费”。
更重要的是降低供应商进入市场、测试需求和建立用户渠道的初始成本。
对于一个正在形成的交易市场而言,供给越丰富,竞争越充分,价格发现就越有效。
③ API 竞价:让价格真正跟随供需变化
传统 API 是固定价目表。
Unexhub希望把它变成动态竞价市场。
当某一时间段算力供应充足,开发者有机会获得更具竞争力的价格。
当需求快速增长,供应商则能够通过市场价格获得更高回报。
这意味着:
价格不再完全由单一厂商决定,而是由供需关系共同形成。
对于开发者而言,平台主打的是按 Token 实际使用量付费,并通过供应商之间的竞争寻找更低的调用成本。
Unexhub目前给出的平台权益中,开发者在符合活动及供应条件的情况下,最低可享受市场价格约 3 折的调用成本。
而供应商侧,则可以通过一键接入渠道、实时结算调用分成等方式,更快将算力转化成收入。
一边降低开发者采购成本,一边提高供应商资金周转效率。
这才是交易市场真正需要形成的双向循环:
更低成本 → 更多开发者 → 更多调用 → 更多供应商 → 更充分竞争 → 更好的价格发现。
AI 视频的成本结构正在被重新计算
MiniMax Design 和 Unexhub 位于 AI 产业链的两个不同位置,却指向了同一个变化。
应用层:降低人的操作成本。
基础设施层:降低机器的算力成本。
最终形成的是一场双重优化:
人花更少时间操作工具,机器用更低成本完成计算。
这也是为什么,AI 行业正在逐渐从“模型参数”和“Token 单价”,转向一个更加现实的问题:
完成一项真正工作的成本是多少?
视频创作者真正关心的,不是一次调用到底消耗了多少 Token。
而是:
我花多少钱,能不能稳定得到一条可以商业使用的视频?
程序员同样如此。
他真正关心的不是 Agent 消耗了多少 Token,而是:
这个任务到底有没有成功完成?
因此,AI 的经济模型正在从:
Token Economics
逐渐走向:
Task Economics。
视频创作正在进入更高的抽象层
过去制作一条商业视频,需要编剧、分镜师、导演、摄影师、剪辑师、配乐师等多个角色共同完成。
AI 并不一定意味着这些角色消失。
真正发生变化的是:
这些专业能力正在被重新封装。
创作者不再需要首先理解复杂的软件流程。
也不一定需要知道如何连接每一个 ComfyUI 节点。
更不需要在多个工具之间反复切换。
你只需要说清楚:
“我想要什么。”
而系统越来越负责:
“应该怎么做。”
这正是 Codex 类智能体带来的核心变化。
视频模型也有了自己的 Codex
所以,MiniMax Design 的意义并不只是又推出了一个 AI 视频工具。
它更像是一次工作方式的升级:
从模型走向 Agent,
从 Prompt 走向 Workflow,
从像素走向语义,
从生成走向生产。
而当 AI 视频真正进入商业生产之后,下一场竞争必然会落到算力经济学上。
MiniMax Design 试图让视频创作变得更简单。
Unexhub 则试图让 AI 算力交易变得更开放、更灵活、更具价格竞争力。
两者结合起来,指向的是一个更大的趋势:
未来的创作者可能不再需要首先理解模型、API 和基础设施。
他们只需要描述最终结果。
Agent 负责组织工作流。
而市场负责寻找完成这项工作的更优算力成本。
视频模型已经拥有了自己的 Codex。接下来真正值得观察的,是这种“语义级创作”能否从令人惊艳的演示,真正进入企业的日常生产线;而一个更加开放的 Token 与算力交易市场,又能否让这种转变拥有足够可持续的经济基础。