多模态开源席卷8月:企业如何通过算力交易中心本地部署 H3,算力成本直降三分之二

On the same day, 16 leading domestic and foreign chip manufacturers and developer communities completed full compatibility adaptation at the Day0 speed, including Moore Threads, Hygon Information, Bili Technology, and Intel Arc Pro B70.


Open-Source Models + Computing Power Trading: The Cubic Decline of AI Inference Costs

On August 3, an underrated industry milestone reshaped the underlying logic of AI commercialization: MiniMax officially open-sourced its new-generation universal full-modal generation system H3. On the same day, 16 leading domestic and foreign chip manufacturers and developer communities completed full compatibility adaptation at the Day0 speed, including Moore Threads, Hygon Information, Bili Technology, and Intel Arc Pro B70.

Moore Threads fully adapted the SGLang MUSA inference framework, MATE, and mudNN high-performance operator library based on its MTT S5000 AI training and inference acceleration card and MUSA software stack in just 3 hours. Multiple mainstream hardware vendors launched multi-GPU deployment solutions simultaneously, forming a comprehensive hardware ecological barrier for H3.

This is far more than a routine model release. As a top-tier universal full-modal system, MiniMax H3 unifies the contextual understanding of text, image, video, and audio, and natively outputs 2K-resolution, 15-second maximum videos with stereo audio. It ranks first globally with an Elo score of 1130 on the Artificial Analysis audio and video editing benchmark. Its ultra-low generation cost of ¥0.8 per second for 2K video has rewritten the industry cost benchmark.

The most valuable breakthrough is not limited to model parameters and pricing, but a structural industry revolution: the thresholds for open-source model usage and high-performance computing power acquisition are collapsing simultaneously. The superposition of the two is driving a cubic decline in AI application inference costs.

Engineering Value of H3 Open-Source: From Third-Party API Calls to Secure Local Deployment

The full H3 system consists of three core modules: H3 Context IR, H3 Base, and H3 Regenerate-2K. Currently, the core H3 Base module is fully open-sourced, supporting flexible deployment and secondary development via mainstream inference frameworks and workflows such as SGLang, vLLM, diffusers, and ComfyUI.

This open-source capability completely reverses the traditional enterprise AI usage model. In the past, enterprises with high data sovereignty and compliance requirements could only rely on closed third-party APIs, facing three major pain points: data cross-domain leakage risks, rigid per-token billing rules, and usage restrictions locked by supplier pricing and rate limits.

With the open-source H3 Base, enterprises can download model weights for free, deploy locally or on private nodes, and perform customized fine-tuning based on exclusive business data. The whole process realizes zero data cross-domain, independent inference control, and one-time investment for long-term repeated use, eliminating continuous model service fees and hidden compliance costs.

However, local deployment still faces a core bottleneck for SMEs: the high threshold of self-built GPU clusters. Purchasing professional computing hardware and building operation and maintenance teams requires huge capital investment ranging from tens of thousands to millions, which is unaffordable for most small and medium-sized teams. This is where the computing power trading revolution comes into play.

Computing Power Trading Logic: Transforming Idle Hardware Assets Into Standard Token Services

The core innovation of the new computing power ecosystem is to decouple computing power from fixed hardware assets and turn it into on-demand, token-metered lightweight services.

The "People's Data · Urban Lexicon Factory" platform takes the lead in realizing compliant circulation of computing power and data resources, adhering to the core principle of "no cross-domain data, localized computing power, and compliant asset circulation". It deploys open-source models on local nodes, encapsulates idle computing resources into standardized token services, and supports enterprise on-demand calls to maximize resource utilization.

Wuxi’s "Token Supermarket" further lowers the threshold for industrial use. It aggregates more than 50 mainstream high-quality domestic and foreign large models including DeepSeek, Kimi, and GLM. Enterprises no longer need to purchase graphics cards or maintain servers, and only pay for the actual token consumption, realizing zero hardware asset investment and flexible cost control.

Unexhub takes this lightweight service logic to the extreme, positioning itself as a professional token aggregation and distribution computing power trading center. It solves the long-standing pain points of traditional computing power transactions through three core mechanisms, reshaping the global AI token service system.

1. Instant Settlement to Eliminate Supplier Capital Pressure

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

2. Zero-Commission Incentives for Early Access

The industry’s universal service commission rate ranges from 10% to 30%. Unexhub launches an aggressive early-stage policy: zero platform commission for settled suppliers. It sacrifices short-term platform revenue to rapidly expand high-quality computing power and model supply, creating a win-win ecosystem for suppliers and developers.

3. API Bidding Mechanism for Market-Driven Pricing

Adopting a stock exchange-like bidding model, Unexhub realizes dynamic floating API prices driven by real-time market supply and demand. During off-peak periods, low prices help individual developers and small teams reduce experimental costs; during peak demand, suppliers gain higher reasonable returns. The pricing right returns to the market, completely solving the opacity and arbitrariness of traditional platform unilateral pricing.

The three mechanisms jointly resolve the two core contradictions of the computing power industry: the cash flow pressure on the supply side and the pricing uncertainty on the demand side, laying a solid foundation for large-scale popularization of AI computing power.

Cubic Cost Decline: The Superposition Effect of Model Open-Source + On-Demand Computing Power

The combination of open-source model dividends, local deployment advantages, and computing power trading mechanisms brings a qualitative leap in cost reduction, achieving a rare cubic-level cost drop in the AI industry rather than linear reduction.

Open-source models eliminate model usage fees: Closed API services require continuous per-token billing, while H3 open-source weights are permanently free for download and deployment, cutting off the largest recurring model service cost.

Local deployment eliminates compliance and variable costs: Data runs in private nodes without cross-domain transmission, avoiding additional compliance audit costs; one-time deployment supports long-term reuse, getting rid of unpredictable incremental billing of closed APIs.

Computing power trading eliminates fixed asset investment: No need for self-built GPU clusters or operation and maintenance teams. Computing power is purchased on demand with dynamic bidding prices, realizing flexible cost scaling with business volume.

Taking multi-modal video generation projects as an example, the traditional closed-source solution bears triple costs: high model API fees, additional data compliance costs, and rigid fixed computing costs, with no negotiation space. The new solution only retainsdynamic bidding computing power fees, achieving a comprehensive cost overhaul.

Objective Boundaries & Future Trends

This revolutionary cost reduction logic has clear phased boundaries. First, H3 Base is currently open-sourced, while H3 Context IR and H3 Regenerate-2K modules are temporarily unavailable for open deployment and only open official API access, so the full workflow still relies on official services. Second, the API bidding mechanism requires sufficient supply and demand liquidity to achieve accurate price discovery. The zero-commission early-stage policy is designed to solve the cold-start problem of ecological liquidity.

Even so, the industry trend is irreversible: open-source models are lowering the technical threshold for AI usage, and computing power trading platforms are lowering the operating threshold for AI landing. The superposition of the two is the real turning point for large-scale commercial implementation of AI applications.

Unexhub: Reshaping the Global AI Token Service Ecosystem

As a leading token aggregation and computing power trading platform, Unexhub has gathered a full range of high-quality AI resources, covering mainstream large model vendors such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as high-quality personalized models provided by individual developers and SMEs.

It adopts a market bidding mechanism + millisecond-level intelligent routing to screen low-cost and high-quality resources for users in real time. Developers and enterprise users enjoy core exclusive rights and interests: the lowest discount is as low as 30% of the official market price, supporting high-intensity batch calls and large-scale commercial use with extremely low costs.

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


开源模型 + 算力交易:实现 AI 推理成本的三次方级下降

8月3日,AI 行业发生了一件极具变革意义、却被多数人低估的关键事件:MiniMax 正式开源新一代通用全模态生成系统 H3。同日,摩尔线程、海光信息、壁砺科技、英特尔锐炫 Pro B70 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,摩尔线程依托 MTT S5000 AI 训推一体智算卡与 MUSA 软件栈,仅耗时3小时,就完成了从 SGLang MUSA 推理框架到 MATE、mudNN 高性能算子库的全链路适配,多家硬件厂商同步落地多卡 GPU 部署方案,为 H3 搭建起完善的硬件生态底座。

这绝非一次普通的模型开源发布。MiniMax H3 是业界顶尖的通用全模态生成系统,可统一解析文本、图像、视频、音频多模态上下文,原生输出最高2K分辨率、最长15秒、自带立体环绕原声的高清视频。在 Artificial Analysis 音视频编辑权威榜单中,H3 以1130的 Elo 评分稳居全球第一,2K 视频生成成本低至0.8元/秒,大幅刷新行业低价标杆。

比参数、价格更核心的,是一场正在重塑行业的结构性变革:开源模型的使用门槛持续走低,算力资源的获取门槛不断下放,二者深度叠加,让 AI 应用的推理成本,告别线性下降,迈入三次方级暴跌的全新阶段。

H3 开源的核心工程价值:从依赖第三方 API,转向自主可控本地部署

MiniMax H3 整体架构分为三大核心模块:H3 Context IR、H3 Base、H3 Regenerate-2K。目前核心的 H3 Base 模块已完全开源,全面兼容 SGLang、vLLM、diffusers、ComfyUI 等主流推理框架与工作流,支持企业灵活部署、二次开发与个性化调优。

这套开源能力,彻底颠覆了传统企业 AI 落地模式。过往,对数据主权、合规性要求极高的企业,只能被动使用闭源第三方 API,长期面临三大痛点:数据必须出域上传、存在泄露风险;全程按 Token 按量计费,无成本自主权;调用速率、定价规则完全被供应商锁定,业务拓展受限。

而 H3 Base 开源后,企业可免费下载模型权重,部署在本地或私有节点,依托自有业务数据完成专属微调,实现数据全程不出域、推理完全自主可控、一次部署长期复用,彻底砍掉持续的模型服务费与合规隐形成本。

但本地规模化落地,始终绕不开中小企业的核心痛点:自建 GPU 集群门槛极高。专业算力硬件采购、服务器搭建、运维团队组建,需要数万至千万级资金投入,绝大多数中小团队难以承担。而算力交易生态的成熟,恰好补齐了这最后一块短板。

算力交易全新逻辑:闲置硬件资产,转化为标准化 Token 服务

新时代算力生态的核心变革,是将算力从「固定资产」解绑,转化为按需取用、Token 计价的轻量化服务,彻底降低 AI 运行门槛。

「人民数据·城市词库工厂」率先落地合规算力流转模式,坚守数据不出域、算力本地化、资产合规流转的核心原则,将开源模型本地化部署,把闲置算力资源封装为标准化 Token 服务,供企业按需调用,最大化盘活闲置算力价值。

无锡「词元超市」进一步简化产业落地路径,聚合 DeepSeek、Kimi、GLM 等50余款海内外主流优质大模型。企业无需采购显卡、无需运维服务器,仅根据实际消耗的 Token 结算费用,实现零硬件资产投入、成本灵活可控。

而 Unexhub 将这套轻量化逻辑推向极致,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,解决传统算力交易的行业顽疾,重塑全球 AI Token 服务格局。

一、即时结算,彻底化解供应商资金压力

传统算力平台普遍采用月结、季结模式,中小算力供应商回款周期长、现金流压力巨大。Unexhub 搭建独立金融清算体系,实现算力供给多少、收益即时到账,让供应商资金压力趋近于零,保障生态供给稳定。

二、早期零佣金激励,让利生态共建者

行业通用算力平台佣金率普遍高达10%-30%。Unexhub 推出极具竞争力的早期政策:入驻供应商可享受最低0平台佣金。平台主动让出短期收益,快速扩充优质算力与模型供给,打造供需双赢的良性生态。

三、API 竞价模式,市场定价替代平台垄断定价

首创「算力交易所」式竞价机制,API 价格随市场供需动态浮动。闲时算力低价普惠,助力个人开发者、中小团队低成本开展技术实验;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,彻底解决传统平台定价不透明、定价僵化的问题。

三大机制相辅相成,完美破解算力行业两大核心矛盾:供给侧的资金周转难题、需求侧的定价不透明痛点,为 AI 算力规模化普及筑牢根基。

三次方级成本下降:开源+本地部署+算力交易的三重叠加红利

开源模型红利、私有化部署优势、市场化算力交易三者叠加,带来的不是线性降本,而是质变级的三次方成本暴跌,彻底重构 AI 项目成本结构。

开源模型,清零模型使用费:告别闭源 API 持续按 Token 扣费模式,H3 模型权重免费下载、永久复用,直接砍掉 AI 应用最大的持续性开销。

本地部署,清零合规与变量隐形成本:数据私有部署、全程不出域,规避数据泄露与合规审核额外开销;一次部署长期复用,摆脱按量付费的不可控增量成本。

算力交易,清零固定资产投入:无需自建算力集群、无需运维硬件设备,算力按需采购、价格动态竞价,成本随业务规模灵活伸缩,无闲置资源浪费。

以多模态视频生成项目为例,传统闭源方案需承担「模型API费+合规服务费+固定算力成本」三重开销,无任何议价空间;全新方案仅保留动态竞价算力成本,实现成本结构的全面优化。

客观边界与行业趋势

这套降本逻辑存在阶段性边界:目前仅 H3 Base 模块开源,H3 Context IR、H3 Regenerate-2K 尚未开放本地部署,完整工作流仍需依托官方 API;同时,API 竞价模式需要足够的供需流动性支撑,平台早期零佣金政策,正是为了解决生态冷启动问题,快速积累供需资源。

但行业趋势已然清晰:开源模型持续降低 AI「使用门槛」,算力交易持续降低 AI「运行门槛」,两者深度融合,正是 AI 应用大规模商业化落地的核心转折点。

Unexhub:重塑全球 Token 逻辑,打造一站式 AI 算力交易枢纽

作为行业领先的 Token 聚合与算力交易平台,Unexhub 全面整合全球优质 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。

平台依托毫秒级智能路由+市场化竞价机制,为用户实时筛选低价、稳定、高质的算力与模型资源。核心权益优势突出:用户调用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。Unexhub 以透明、高效、低价的核心优势,为全球开发者与企业,搭建起全新的 AI 资源流通生态。