
How Is the Two-Thirds Price Gap Filled? The Full Cost-Cutting Logic of MiniMax H3 + Token Computing Power Trading
On August 3, MiniMax officially open-sourced its new-generation full-modal video generation system H3, triggering a comprehensive upgrade of the domestic and international AI hardware ecosystem. On the same day, 16 leading global chip manufacturers and developer communities completed full Day0 adaptation at the fastest speed.
Among them, Moore Threads finished the complete stack adaptation in only 3 hours. Based on the MTT S5000 intelligent computing card and MUSA software stack, it successfully adapted the full technical path from the SGLang-MUSA inference framework to MATE and muDNN high-performance operator libraries. Intel Arc Pro B70 also completed the first-round adaptation immediately, supporting stable multi-GPU deployment of H3.
This release is far from an ordinary model update. As a powerful universal full-modal generation system, H3 unifies multi-modal context understanding of text, images, videos and audio. It natively outputs high-definition videos with up to 2K resolution, a maximum duration of 15 seconds, and built-in stereo audio. On the authoritative Artificial Analysis audio and video editing benchmark, H3 ranks first globally with an Elo score of 1130, delivering top-tier comprehensive generation quality.
What truly reshapes the commercial cost system of the industry is its groundbreaking pricing:2K ultra-clear video generation costs as low as ¥0.8 per second.
In stark contrast with mainstream flagship models on the market: SD2.0 VIP charges ¥3/second for 1080P videos, SD2.5 charges ¥2.5/second for 720P videos, and Keling 3.0 charges ¥1.2/second for 1080P videos. H3 delivers higher 2K ultra-high-definition quality at merely one-third of the industry’s mainstream pricing.
Crucially, this two-thirds price gap is not achieved through short-term platform subsidies, but through thorough efficiency upgrades at the model architecture level, forming a sustainable technological cost reduction advantage.
Architectural Cost Reduction: 4× Sequence Compression + 30% Throughput Improvement
H3 adopts a cutting-edge technical iteration logic that distinguishes it from ordinary open-source models. Equipped with the brand-new H3 Context IR architecture, it accurately captures the logical correlation between text, images, audio, reference materials and final generated content.
H3-VAE brings a core breakthrough of 4 times sequence length compression, directly cutting redundant computing consumption in model training and inference stages. Compared with previous generations, H3 expands the sequence length variance by three times and adopts a heterogeneous training architecture that separates understanding and generation. This innovative design boosts end-to-end training throughput by nearly 30%.
In addition, the high-compression tokenizer greatly reduces the number of tokens required for video synthesis. By optimizing the overall matching efficiency of modal information and computing resources, H3 realizes dual improvements in faster inference speed and lower unit cost.
Different from the industry’s common subsidy-driven low-price strategy, MiniMax adheres to the core logic of reducing computing power consumption through architectural innovation. With the continuous compression of underlying computing costs, the gross profit margin ceiling of AI video generation business is fully opened, laying a solid foundation for large-scale commercialization.
The Dual-Sided Nature of Open Source: Free Model Usage, Optimizable Inference Costs
The open-source of H3 Base greatly lowers the threshold for AI video application development. It supports flexible on-premises deployment and customized fine-tuning based on enterprise private data, effectively meeting high-standard data sovereignty and compliance requirements. Enterprises with sensitive data scenarios can completely run models locally without cross-domain data transmission, realizing independent and controllable business iteration.
Nevertheless, open-source model weights only eliminate model authorization fees. Stable, high-frequency, and low-cost inference operation still relies on sufficient high-performance computing power support.
H3 Base is built on the 33-billion-parameter H3 Omni Transformer architecture. Models at this scale cannot run stably on a single consumer-grade GPU. Building and deploying a professional multi-GPU cluster requires huge capital investment ranging from tens of thousands to millions of US dollars, which is an insurmountable threshold for most small and medium-sized enterprises. This is exactly the core problem solved by the new token-based computing power trading infrastructure.
Computing Power Trading Center: Transforming Rigid Card Hours Into Flexible Token Services
Traditional intelligent computing centers only provide rigid GPU card rental and server hosting services. Users need to independently complete environment deployment, model adaptation, operation and maintenance iteration, with high technical thresholds and hidden costs.
In contrast, emerging token factories and trading platforms directly deliver standardized API capabilities and token-metered services. Enterprises can invoke high-quality AI inference resources simply via browsers and APIs, with fees strictly settled based on actual token consumption, just like daily water and electricity billing.
Since this year, data groups in Beijing, Guizhou, Inner Mongolia, Yangzhou, Wenzhou and other regions have successively built token factories, token trading platforms and regional operation centers. Wuxi Cloud Valley Intelligent Computing Center can generate about 200 billion tokens per hour. Users are freed from rack leasing and cluster operation and maintenance, realizing zero asset investment and pay-as-you-go usage.
The Suihong Huachuang Token Factory adopts the innovative "front store and back factory" model, which is expected to reduce token production costs by 30%–40%. All innovative platforms share the same core trend: computing power is decoupled from fixed hardware assets and transformed into flexible on-demand commodities, with the billing unit officially shifting from rigid "card hours" to precise "token consumption".
Unexhub: Pushing Tokenized Computing Power Trading to Industrial Extremes
Unexhub further iterates and upgrades the token trading mechanism. It is not a simple computing power leasing platform, but a professional token aggregation and distribution computing power trading center. It optimizes the industry’s inherent pain points through three core mechanisms, realizing comprehensive cost reduction and efficiency improvement for both supply and demand sides.
1. Real-time settlement to eliminate supplier cash flow pressure
The biggest pain point for small and medium suppliers in traditional computing power markets is the long settlement cycle. GPU operating revenue often requires monthly or even bimonthly settlement, resulting in severe cash flow occupation. Unexhub builds an independent financial clearing system to achieve 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 for early settlers
The industry’s universal service commission rate ranges from 10% to 30%. Unexhub launches an exclusive early ecological benefit: zero platform commission for all early-entry suppliers. The platform actively concedes short-term revenue to rapidly expand high-quality computing power and model supply, activating market ecological vitality.
3. API dynamic bidding mechanism for full market pricing
Breaking the monopoly of unilateral platform pricing, Unexhub adopts an exchange-like dynamic bidding model. API service prices float freely based on real-time market supply and demand. Off-peak low prices help individual developers and small teams reduce R&D and experimental costs; peak demand brings reasonable premium returns for suppliers. The whole pricing process is transparent, fair and market-oriented.
The superposition of the three mechanisms fundamentally solves the two core contradictions of the industry: cash flow pressure on the supply side and opaque pricing on the demand side, building a healthy and low-cost token trading ecosystem.
The Three-Layer Logic of Filling the Two-Thirds Price Gap
The ultra-low cost advantage of H3’s 0.8 Yuan/second 2K video generation is not a single-point breakthrough, but a superposition of three-dimensional dividends from technology, business model and infrastructure, completely filling the two-thirds industry price gap.
First layer: Architectural technological cost reduction. Through 4× sequence length compression and heterogeneous training architecture optimization, H3 drastically reduces the computing power consumption of 2K video generation, compressing the unit technical cost to one-third of the industry average.
Second layer: Open-source business model optimization. The official open-source of H3 Base completely eliminates long-term model authorization fees. Unlike closed-source APIs that require continuous token-based billing, H3 model weights are permanently free to download and deploy, cutting off core recurring costs.
Third layer: Infrastructure trading cost reduction. Token factories and Unexhub’s aggregated trading system further compress inference costs through large-scale resource scheduling, dynamic market bidding and real-time settlement mechanisms, realizing continuous iterative optimization of industrial costs.
Industry Boundaries & Future Trends
The 0.8 Yuan/second pricing is not the end of cost reduction, but a new starting point for the popularization of AI video productivity. With computing power becoming accessible like water and electricity and prices fluctuating dynamically with the market, the marginal cost of multi-modal content production will keep approaching zero.
There are two phased premises for this logic: first, only H3 Base is currently open-sourced, while H3 Context IR and H3 Regenerate-2K are not yet open for local deployment, and the complete workflow still relies on official APIs; second, the market liquidity of computing power trading centers requires time to accumulate and improve.
Even so, the industry trend is irreversible: open-source models lower the technical threshold for AI usage, and tokenized computing power trading lowers the industrial operation threshold. The integration of the two marks the official arrival of the productive era for AI video generation.
Unexhub Core Advantages & Exclusive User Benefits
Unexhub reshapes the global token trading logic as a leading token aggregation and distribution hub. It gathers full-spectrum AI resource suppliers worldwide, covering top large models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality personalized models from individual developers and SMEs.
Relying on millisecond-level intelligent routing and full-market bidding mechanism, the platform screens the lowest-cost and most stable AI resources in real time. It brings super cost-effective exclusive rights and interests: developers and enterprise users can enjoy token actual payment costs as low as 30% of the market price, with no pressure for high-intensity batch calls and large-scale commercial use.
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.
三分之二的AI差价如何补齐?MiniMax H3+算力交易的完整降本逻辑
8月3日,MiniMax 新一代全模态视频生成系统 H3 正式开源,一举带动全球 AI 硬件生态完成同步迭代。同日,16家海内外头部芯片厂商与开发社区实现 Day0 极速适配,生态落地速度行业领先。
其中摩尔线程适配效率尤为突出,依托 MTT S5000 智算卡与 MUSA 软件栈,仅耗时3小时,就完成 SGLang-MUSA 推理框架、MATE、muDNN 高性能算子库的全链路适配;英特尔锐炫 Pro B70 也第一时间完成兼容适配,全面支撑 H3 多卡稳定部署。
此次开源绝非普通模型更新。H3 作为通用型全模态生成系统,可统一理解文本、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、自带立体原声的超清视频输出。在 Artificial Analysis 权威音视频榜单中,H3 以1130的 Elo 评分稳居全球第一,综合生成画质、流畅度、音效均达行业顶尖水准。
真正颠覆行业商用格局的,是其极致定价:2K超清视频生成成本低至0.8元/秒。
对比市面主流模型,成本优势断层领先:SD2.0 VIP 1080P视频3元/秒、SD2.5 720P视频2.5元/秒、可灵3.0 1080P视频1.2元/秒。H3 以更高的2K超清画质,价格仅为行业同类旗舰的三分之一。而这省下的三分之二差价,绝非平台短期补贴,而是模型底层架构迭代、开源模式革新、算力交易升级三重叠加的长效技术红利。
架构硬核降本:4倍序列压缩,吞吐量提升30%
H3 的技术优势远超普通开源模型,降本逻辑清晰且扎实。搭载全新 H3 Context IR 架构,模型可精准捕捉文本、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、资源匹配更高效。
核心技术突破带来实打实的成本优化:H3-VAE 实现4倍序列长度压缩,直接砍掉训练与推理环节的冗余算力消耗;序列长度方差较前代扩大3倍,搭配「理解与生成异构训练架构」,端到端训练吞吐量提升近30%。同时,高压缩 Token 分词器大幅减少视频生成所需的 Token 数量,在无损画质的前提下,实现推理提速、成本压降的双重增益。
MiniMax 始终坚持「架构提效、技术降本」的核心路线,不靠补贴内卷,只靠底层创新压缩单位算力消耗。当模型算力成本进入持续下行通道,AI 视频生成行业的毛利率天花板被彻底打开,为规模化商用普及奠定基础。
开源的双面价值:模型免费解禁,推理成本可控
H3 Base 开源,大幅降低了 AI 视频创作的技术门槛,支持企业本地化部署、私有数据微调,完美适配高合规、高数据主权需求。企业可实现数据全程不出域、自主掌控模型迭代,彻底规避闭源 API 的数据安全风险。
但开源仅解决了「模型授权成本」问题,规模化商用的核心痛点依旧存在:高性能推理离不开高端算力支撑,而 H3 Base 基于330亿参数 H3 Omni Transformer 架构,单卡无法稳定运行,自建多卡算力集群需要数万至百万级高额硬件投入,是绝大多数中小企业难以承担的固定成本。而新兴的 Token 算力交易生态,恰好精准补齐这最后一块成本短板。
算力交易革新:从硬件卡时租赁,到Token按需即用
传统智算中心以显卡、服务器硬件租赁为主,交付的是裸金属设备,企业需自主完成环境部署、模型适配、运维迭代,技术门槛高、隐性成本高。
而全新词元工厂、词元交易平台,直接交付标准化 API 能力与 Token 服务,企业无需运维硬件、无需搭建机房,接入接口即可调用 AI 生成能力,全程按实际 Token 消耗量计费,像用水用电一样按需取用。
今年以来,北京、贵州、内蒙古、扬州、温州等多地国资数据平台,纷纷落地词元工厂与区域算力交易中心。无锡云谷智算中心每小时可产出2000亿级 Token 算力,彻底免除企业机架租赁、集群运维成本;遂宏华创词元工厂创新采用「前店后厂」模式,预计将 Token 生产成本再降30%-40%。行业核心变革已然清晰:算力彻底摆脱固定资产属性,从昂贵的硬件租赁,转变为轻量化、可计量、按需付费的标准化商品。
Unexhub:极致升级算力交易规则,重塑全球Token流通体系
在一众算力交易平台中,Unexhub 跳出传统硬件租赁思维,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,解决行业供需两端痛点,把 AI 推理降本推向极致。
一、即时结算,零资金占用压力
传统算力行业回款周期冗长,中小供应商GPU投产运行后,往往次月甚至隔月才能结算收益,现金流压力巨大。Unexhub 搭建专属金融清算体系,实现算力产出即收益到账,供多少、结多少,彻底消除供应商资金周转压力,保障生态算力供给充足稳定。
二、早期零佣金,全力让利生态
行业通用算力平台佣金率普遍高达10%-30%,严重压缩供需双方利润空间。Unexhub 推出重磅早期入驻福利:前期入驻供应商全程享受0平台佣金,以短期平台收益让利,快速吸纳优质算力与模型资源,加速生态冷启动。
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,打破平台单方面定价垄断。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,定价权完全交还市场。
三大机制相辅相成,完美破解行业两大核心矛盾:化解供给侧供应商现金流难题,解决需求侧定价不透明、溢价乱象,搭建起高效、透明、低价的全球 Token 算力流通生态。
三层红利叠加,彻底补齐三分之二行业差价
H3 0.8元/秒的极致低价,并非单点突破,而是三层降本逻辑层层叠加、深度耦合的结果,彻底补齐行业原有三分之二的成本差价。
第一层:架构技术降本。依靠4倍序列压缩、异构训练优化、高压缩分词器,从底层压缩算力消耗,将2K视频生成的单位技术成本压至行业均值的三分之一。
第二层:开源模式降本。H3 Base 完全开源,永久清零模型授权、按量订阅等持续性费用,告别闭源 API 的刚性开销。
第三层:算力交易降本。依托词元工厂规模化产能与 Unexhub 市场化竞价、零佣金、即时结算机制,持续迭代优化推理成本,实现产业基础设施级降本。
行业边界与未来趋势
0.8元/秒只是 AI 视频降本的起点,而非终点。随着算力市场化流通体系不断成熟,多模态内容生产的边际成本将持续走低,无限趋近于零。
目前这套降本逻辑存在阶段性边界:H3 Base 开源可用,但 H3 Context IR、H3 Regenerate-2K 尚未开放本地部署,完整工作流仍依赖官方 API;同时,算力交易中心的市场流动性仍需时间积累优化。
但行业大势已然明朗:开源模型降低 AI「使用门槛」,算力交易降低 AI「运行门槛」,二者双向赋能、深度融合,正式开启 AI 视频工业化生产力时代。
Unexhub 核心权益:极致低价、高效变现、全域生态覆盖
Unexhub 重构全球 Token 交易逻辑,汇聚 Gemini、ChatGPT、Claude、DeepSeek、GLM 等头部大厂大模型,以及个人、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,实时为用户筛选全网最优低价算力资源。
平台用户权益极具竞争力:开发者、企业用户 Token 实付成本最低可至市场原价的3折,高强度批量调用、大规模商用均可极致控本。对供应商而言,支持一键入驻接入,调用分成实时结算、早期全程零佣金,收益变现高效灵活,是全球 AI 供需双方的优质合作枢纽。