MiniMax H3 开源后,企业算力成本怎么算?算力交易中心给出 0.8 元/秒的 2K 音画答案

What truly reshapes the commercial landscape of AI video generation is its groundbreaking pricing: 2K ultra-clear video generation costs as low as 0.8 Yuan per second, setting a new industry low for high-quality multi-modal content production.


0.8 Yuan/Second for 2K AI Video: The Real Cost Breakdown Behind MiniMax H3

On August 3, MiniMax officially open-sourced H3, its new-generation universal full-modal video generation model. On the same day, 16 leading domestic and international chip manufacturers and developer communities completed full Day0 adaptation, including Huawei Ascend, Moore Threads, Muxi, Hygon Information, Kunlunxin, Bili Technology, AMD, and Intel. Among them, Moore Threads achieved a comprehensive adaptation of the full technical stack within only 3 hours, covering the SGLang-MUSA inference framework, MATE, and muDNN high-performance operator libraries based on its MTT S5000 accelerator card and MUSA software stack.

This open-source release is far from an ordinary model update. As a powerful universal full-modal generation system, H3 can comprehensively understand multi-modal context composed of text, images, videos and audio. It natively supports generating ultra-clear videos with up to 2K resolution, a maximum duration of 15 seconds, and built-in stereo audio. It ranks first globally with an Elo score of 1130 on the authoritative Artificial Analysis audio and video editing benchmark list.

What truly reshapes the commercial landscape of AI video generation is its groundbreaking pricing: 2K ultra-clear video generation costs as low as 0.8 Yuan per second, setting a new industry low for high-quality multi-modal content production.

Compared with mainstream flagship models in the industry, H3’s cost advantage is extremely prominent: SD2.0 VIP charges 3 Yuan/second for 1080P video, SD2.5 charges 2.5 Yuan/second for 720P video, and Lingke 3.0 charges 1.2 Yuan/second for 1080P video. H3 delivers superior 2K ultra-high-definition output quality at only one-third of the cost of competing flagship models.

Notably, the ultra-low price of 0.8 Yuan/second is not supported by financial subsidies, but achieved entirely through revolutionary architectural efficiency upgrades and technological iteration.

H3 adopts the cutting-edge H3 Context IR model architecture, which accurately identifies the logical correlation between text, image, audio, reference materials and expected generation results. Equipped with H3-VAE technology, it realizes 4x sequence length compression, greatly reducing hardware resource consumption and lowering overall training and inference costs.

Compared with previous generations, H3 expands the sequence length variance by three times. It adopts a heterogeneous training architecture that separates understanding and generation, boosting end-to-end training throughput by nearly 30%. Meanwhile, the high-compression tokenizer drastically cuts the number of tokens required for video generation, realizing dual improvements in inference speed and cost control.

Different from the industry’s common subsidy-driven low-price strategy, MiniMax adheres to the core logic of improving performance and reducing costs through architectural innovation. Technological iteration continuously compresses unit computing power consumption, fundamentally breaking the gross profit margin ceiling of traditional AI video generation business and opening up large-scale commercial space for enterprises.

Nevertheless, open-sourcing is only the first step of industrial popularization. The model weights are now free to download for all developers and enterprises, but stable and low-cost inference operation relies on sufficient computing power support. The emerging token-based computing power trading ecosystem perfectly solves the pain point of "difficult and expensive computing power acquisition".

Computing Power Trading: Transforming Rigid Hardware Assets Into Flexible Token Services

On the day of H3’s open-source release, cloud inference platforms including RunningHub completed synchronous adaptation. Creators no longer need to purchase expensive high-end graphics cards and build local deployment environments. They can call the H3 model with zero threshold on cloud platforms to realize high-speed rendering of 2K ultra-clear videos.

Behind this convenient cloud inference capability is a mature and improved new computing power infrastructure system. The National Data Bureau designates 2026 as the "Year of Data Element Value Release", clearly proposing to explore the value system and trading mechanism of token elements and promote compliant and efficient circulation of computing power and data resources.

Adhering to the principle of"no cross-domain data transmission, localized computing power, and compliant asset circulation", the "People’s Data · Urban Lexicon Factory" encapsulates idle social computing resources into standardized token services, providing stable on-demand access for enterprises and effectively activating idle resource value.

Launched in May 2026, Wuxi’s "Token Supermarket" aggregates more than 13,000 PFLOPS of intelligent computing power. It transforms high-cost centralized computing power into lightweight token-metered AI services. Enterprises can access mainstream open-source models such as Qwen, DeepSeek, MiniMax and ChatGLM on demand without building computer rooms, purchasing hardware or setting up algorithm teams, just like using water and electricity.

Guizhou Mobile’s "Token Factory" has fully deployed mainstream domestic open-source model resources, with comprehensive token costs more than 20% lower than industry average platforms. The Southeast Token Operation Center is compatible with over 100 mainstream open-source, closed-source and third-party token services including DeepSeek, Qwen and GLM, enabling one-stop access and out-of-the-box usage for enterprises.

All innovative platforms share the same core logic: computing power is decoupled from fixed hardware assets and transformed into flexible on-demand commodities, with the billing unit shifting from rigid "card hours" to precise "token consumption".

Unexhub: Pushing Token-Based Computing Power Trading to the Industry Extreme

Unexhub further optimizes and iterates the token trading mechanism, positioning itself as a professional token aggregation and distribution computing power trading center, rather than an ordinary computing power platform. It solves many long-standing pain points in the traditional computing power industry through three core mechanisms, bringing users ultra-low-cost and high-stability AI inference services.

1. Real-time settlement to eliminate supplier cash flow pressure

Traditional computing power platforms adopt monthly or quarterly settlement cycles, leaving small and medium suppliers facing long fund recovery cycles and huge cash flow pressure. Unexhub builds an independent financial clearing system to realize real-time settlement for all valid computing power output. Suppliers get paid immediately after providing services, with nearly zero capital pressure.

2. Zero-commission incentive policy for early settlers

The industry’s universal service commission rate ranges from 10% to 30%. Unexhub launches a generous early-stage benefit: zero platform commission for settled suppliers. The platform actively concedes short-term revenue to rapidly expand high-quality computing power and model supply, building a sufficient and stable resource pool.

3. API bidding mechanism for market-oriented pricing

Initiating an exchange-like dynamic bidding model, Unexhub enables API prices to float freely based on real-time market supply and demand. Prices drop during off-peak periods, allowing individual developers and small teams to conduct experiments and develop projects at ultra-low costs; prices rise moderately during peak demand, bringing reasonable profit returns to suppliers. Pricing is completely dominated by the market, abandoning the arbitrary unilateral pricing of traditional platforms.

Superimposed Cost Reduction: The Exponential Drop Logic of AI Video Generation Cost

The combination of H3’s architectural efficiency dividends and Unexhub’s token trading mechanism completely reconstructs the cost structure of AI video inference, realizing exponential cost compression rather than linear reduction.

In the past, multi-modal video generation projects relying on closed-source APIs faced triple high costs: expensive per-second API fees, additional data compliance review costs, and locked computing power suppliers with no bargaining space.

The new generation solution completely optimizes the cost structure: H3 model weights are permanently free to download and deploy; enterprises can access low-cost computing power nodes through token supermarkets and token factories; inference computing power is purchased at market bidding prices. The cost composition is simplified from "model licensing fee + computing power rental fee + compliance cost" to pure dynamic bidding computing power fee.

H3’s architectural innovation has compressed the industry’s average 2K video generation cost to 0.8 Yuan/second. Unexhub’s token trading mechanism further reduces the operating cost ceiling through real-time settlement, zero-commission supply incentives and market bidding pricing. The dual superposition of model technological progress and computing power mechanism innovation brings an exponential decline in the total cost of AI video generation.

Industry Premises & Future Trends

This revolutionary cost reduction logic has clear phased premises. First, only the H3 Base module is currently open-sourced, while H3 Context IR and H3 Regenerate-2K are not yet open for local deployment, and the complete generation workflow still relies on official API services. Second, the liquidity of the computing power trading market needs time to accumulate and improve.

Even so, the industry development trend is irreversible: open-source models continuously lower the technical threshold for AI usage, and token-based computing power trading continuously lowers the operating threshold for AI landing. The integration of the two marks the official entry of AI video generation into the large-scale industrial productivity era.

Unexhub Core Advantages & Exclusive Benefits

As a global leading token aggregation and computing power trading center, Unexhub gathers full-scale AI resource suppliers, covering international top large models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality personalized models provided by individual developers and SMEs.

Relying on millisecond-level intelligent routing + full-market bidding mechanism, it screens the lowest-cost and most stable resources for users in real time. It brings super cost-effective exclusive rights and interests for global developers and enterprise users: the actual token payment cost is as low as 30% of the official market price, with no pressure for high-intensity batch calls and large-scale commercial production.

For suppliers, one-click platform access is supported, with real-time settlement of call sharing revenue and zero early commission, achieving efficient and flexible profit realization. Unexhub builds a transparent, efficient and low-cost global AI resource circulation ecosystem for both supply and demand sides.


2K视频0.8元/秒,AI视频生成的低成本账单究竟怎么算?

8月3日,MiniMax 正式开源新一代通用全模态视频生成模型 H3。同日,华为昇腾、摩尔线程、沐曦、海光信息、昆仑芯、天数智芯、壁仞科技、AMD、Intel 等16家海内外头部芯片厂商与开发社区同步完成 Day0 极速适配。其中摩尔线程表现尤为亮眼,依托 MTT S5000 智算卡与 MUSA 软件栈,仅耗时3小时,就完成了 SGLang-MUSA 推理框架、MATE、muDNN 高性能算子库的全栈完整适配,为 H3 落地搭建了扎实的硬件生态底座。

此次开源绝非普通的模型迭代,而是重塑 AI 视频行业成本格局的关键动作。H3 作为通用型全模态生成系统,可统一解析文本、图像、视频、音频多模态上下文,原生支持最高2K分辨率、最长15秒、自带立体环绕原声的超清视频生成。在权威榜单 Artificial Analysis 音视频编辑评测中,H3 以1130的 Elo 评分稳居全球综合排名第一。

真正让企业实现规模化商用盈利的,是 H3 颠覆性的定价能力:2K超清视频生成成本低至0.8元/秒,彻底打破行业高价壁垒。

横向对比市面主流旗舰模型,H3 的成本优势碾压同级产品:SD2.0 VIP 1080P视频定价3元/秒,SD2.5 720P视频定价2.5元/秒,可灵3.0 1080P视频定价1.2元/秒。H3 以远超竞品的2K超清画质,价格仅为行业同类旗舰模型的三分之一,性价比优势断层领先。

值得重点说明的是,0.8元/秒的极致低价绝非平台补贴堆砌而来,而是依托底层架构革新与技术效率升级,从根源压缩推理成本实现的可持续低价。

H3 搭载全新升级的 H3 Context IR 模型架构,可精准识别文本、图像、音频、参考素材与生成结果之间的深层关联,推理逻辑更精准高效。搭配 H3-VAE 技术,实现4倍序列长度压缩,大幅降低训练与推理阶段的硬件资源消耗。

同时,模型序列长度方差较前代扩大3倍,采用「理解与生成异构训练架构」,端到端训练吞吐量提升近30%。通过高压缩率 tokenizer 大幅减少视频生成所需的 token 数量,在保证画质无损的前提下,实现速度提质、成本压降的双重突破。

区别于行业普遍的补贴换低价模式,MiniMax 坚持「技术提效、架构降本」的核心路线,以技术迭代持续压缩单位算力消耗,从根源上重构 AI 视频生成的成本结构,彻底打开企业商用的毛利率天花板,为行业规模化落地创造条件。

当然,模型开源只是普惠第一步。目前 H3 权重已免费对外开放下载,彻底清零模型授权成本,但稳定、低成本的规模化推理,仍依赖充足优质的算力支撑。而日趋成熟的Token算力交易生态,恰好补齐了中小企业算力贵、算力难寻的核心短板。

算力交易革新:固定资产算力,变身按需取用的轻量化服务

H3 开源上线当日,RunningHub 等主流云推理平台同步完成适配。创作者无需高价采购专业显卡、搭建本地部署环境,零门槛即可在线调用 H3 模型,云端高速渲染生成2K超清视频,大幅降低个人创作者与中小团队的落地门槛。

便捷云端部署的背后,是国内日趋完善的新型算力基础设施体系。国家数据局将2026年定为「数据要素价值释放年」,明确探索词元要素价值体系与交易机制,推动算力、数据资源合规高效流转。

「人民数据·城市词库工厂」坚守数据不出域、算力本地化、资产合规流转核心原则,盘活社会闲置算力资源,封装为标准化 Token 服务,为企业提供稳定、安全、按需适配的算力支撑,最大化释放闲置资源价值。

2026年5月试运营上线的无锡「词元超市」,已汇聚超13000 PFLOPS 智能算力,将昂贵的集中式专属算力,转化为 Token 计价的轻量化普惠 AI 服务。企业无需自建机房、采购硬件、组建算法运维团队,即可像用水用电一样,按需调用 Qwen、DeepSeek、MiniMax、ChatGLM、Kimi 等主流开源模型能力。

贵州移动「词元工厂」全面布局主流开源模型生态,综合 Token 使用成本较行业通用平台降低20%以上;东南 Token 运营中心兼容 DeepSeek、Qwen、GLM 等100余款开源、闭源主流模型,企业一次对接即可全量开箱即用,大幅降低接入成本与运维成本。

所有算力平台的革新逻辑高度统一:算力正在摆脱「高价固定资产」属性,转变为灵活可控的按需商品,计费模式也从僵化的「显卡时长计费」,升级为精准透明的「Token 按量计费」。

Unexhub:极致迭代Token交易机制,重塑全球AI算力定价体系

在一众算力交易平台中,Unexhub 跳出传统算力租赁的单一模式,精准定位为Token聚合分发型算力交易中心,通过三大核心机制破解行业痛点,将 AI 算力降本逻辑推向极致。

一、即时结算机制,彻底化解供应商现金流压力

传统算力行业普遍存在回款周期长的痛点,GPU 算力投产运行后,往往需要次月甚至隔月结算,中小供应商现金流压力巨大。Unexhub 搭建专属金融清算体系,实现算力产出多少、收益即时到账,零账期、无积压,彻底释放供应商资金压力,保障算力供给稳定充足。

二、早期零佣金红利,让利全生态参与者

行业主流算力平台佣金率普遍高达10%-30%,大幅压缩供需双方利润空间。Unexhub 推出重磅早期扶持政策:入驻供应商可享受最低0平台佣金。平台主动舍弃短期收益,全力吸纳优质算力与模型资源,快速完善生态供给。

三、API市场化竞价,价格透明无垄断

首创「算力交易所」动态竞价模式,彻底颠覆平台单方面定价的行业规则。API 调用价格随市场供需实时浮动:闲时算力低价普惠,助力开发者低成本试错、项目研发;忙时需求上涨,供应商可获得合理溢价收益。全程市场化定价,透明公正,杜绝溢价垄断。

双重降本叠加,实现AI视频成本指数级压缩

H3 架构技术革新 + Unexhub Token 交易机制,双重优势叠加,彻底重构 AI 视频推理成本结构,实现从线性降本到指数级降本的跨越。

过往依托闭源 API 做多模态视频项目,成本压力贯穿全程:单秒调用费用高达数美元,数据上传云端产生额外合规审核成本,算力被单一供应商锁定,无任何议价空间,规模化商用盈利难度极大。

全新落地模式彻底优化成本链路:H3 模型免费开源、永久复用;依托词元超市、词元工厂搭建算力节点;通过 Unexhub 竞价市场获取动态低价算力。成本结构从「模型授权费+算力租赁费+合规服务费」三重开销,精简为单一动态竞价算力成本。

H3 以技术革新将2K视频生成的行业底价压至0.8元/秒,而 Unexhub 通过即时结算、零佣金供给、市场化竞价三大机制,持续压低算力运行底价。模型技术红利与算力机制红利双向叠加,让 AI 视频生成总成本实现指数级下跌,彻底解锁大规模商用盈利空间。

行业边界与未来趋势

目前这套低成本落地逻辑存在阶段性边界:仅 H3 Base 模块完成开源,H3 Context IR、H3 Regenerate-2K 尚未开放本地部署,完整视频生成工作流仍需依托官方 API;同时,算力交易市场的供需流动性仍需时间积累优化。

但行业趋势已然清晰:开源模型持续降低 AI「使用门槛」,Token 算力交易持续降低 AI「运行门槛」,二者深度融合,标志着 AI 视频创作正式告别小众试用阶段,全面迈入工业化生产力时代。

Unexhub 核心权益:低价、高效、全生态覆盖

Unexhub 已汇聚全球全量级 AI 资源,涵盖 Gemini、ChatGPT、Claude、DeepSeek、GLM 等头部大厂大模型,同时吸纳个人开发者、中小企业的优质特色模型,搭建海量供需竞价池。依托毫秒级智能路由技术,实时为用户筛选最优低价算力资源。

平台核心权益极具竞争力:开发者与企业用户 Token 实付成本最低可至官方市场价的3折,高强度批量调用、大规模商用场景也能极致控本、无资金压力。对算力供应商而言,支持一键入驻接入,调用分成实时结算、早期零佣金,收益变现高效灵活,是供需双方共赢的全球 AI 资源流通枢纽。