多模态开源遇上国产 GPU 适配潮:算力交易中心成 2K 音画生成的算力调度枢纽

On the exact day of release, a total of 16 mainstream ecological partners completed full Day0 adaptation, covering domestic and international chip manufacturers, developer communities, cloud inference platforms, and core inference frameworks.


16 Chip Vendors Complete H3 Day0 Adaptation: How Computing Power Trading Centers Integrate Ascend, Muxi and Moore Threads

On August 3, MiniMax officially open-sourced its flagship full-modal generation model H3, triggering an all-round upgrade of the global AI hardware and software ecosystem. On the exact day of release, a total of 16 mainstream ecological partners completed full Day0 adaptation, covering domestic and international chip manufacturers, developer communities, cloud inference platforms, and core inference frameworks.

The adaptation lineup includes 7 leading domestic chip vendors: Huawei Ascend, Moore Threads, Muxi, Hygon Information, Kunlunxin, KUNLUN, and Biren Technology, together with international giants AMD and Intel. It also covers mainstream developer communities and cloud platforms such as Hugging Face, ModelScope, ComfyUI, RunningHub, and fal, as well as core inference frameworks including vLLM-Omni and SGLang.

This synchronous full-scale adaptation is no coincidence. The Day0 seamless compatibility means the brand-new H3 model can be deployed and run stably on all mainstream platforms immediately upon release, completely breaking the past industry situation where new models required months of waiting for hardware adaptation.

Multiple vendors delivered outstanding adaptation efficiency. Muxi took the lead in completing Day0 adaptation for the Xiyun C-series GPU. Moore Threads finished the full technical adaptation path from the SGLang-MUSA inference framework to MATE and muDNN high-performance operator libraries in only 3 hours, relying on its MTT S5000 AI training and inference integrated computing card and MUSA software stack.

For the first time, domestic high-performance chips have fully kept pace with the iteration rhythm of global cutting-edge AI models, realizing a qualitative leap from "months of delayed adaptation" to "same-day full deployment".

H3’s industry-wide appeal stems from its leading comprehensive strength in both performance and commercial value. As a powerful universal full-modal generation system, H3 unifies multi-modal context understanding of text, images, videos and audio, and natively generates high-definition 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 benchmark, surpassing mainstream video models such as Gemini Omni Flash and Xiang 2.7.

Beyond top-tier performance, it is the disruptive commercial cost advantage that drives full vendors’ synchronous Day0 layout. H3’s 2K ultra-clear video generation cost is as low as ¥0.8 per second, merely one-third of the price of mainstream flagship models in the industry. This substantial cost reduction is achieved through systematic architectural optimization: a high-compression tokenizer greatly cuts token consumption during video generation, while heterogeneous training, intelligent load balancing and GPU utilization optimization fully squeeze redundant computing overhead, forming a sustainable low-cost technical barrier.

The superposition of open-source freedom, ultra-low marginal cost and ecological expansion makes H3 adaptation a forward-looking strategic layout for chip vendors targeting the upcoming explosive multi-modal inference market, rather than a meaningless technical public welfare project.

Core Industry Pain Point: Fragmented Heterogeneous Hardware Ecology

Despite the comprehensive synchronous adaptation of 16 vendors, the underlying hardware ecology still faces fragmentation problems. Ascend, Muxi, Moore Threads and other chip brands have independent toolchains, deployment environments and parameter tuning systems. For developers, deploying and running H3 efficiently used to involve cumbersome processes: selecting hardware brands in advance, building exclusive operating environments, and repeatedly debugging adaptive parameters, with high technical thresholds and time costs.

Computing Power Trading Center: The Core Link Connecting Fragmented Hardware Resources

Token-based computing power trading centers perfectly solve the problem of heterogeneous hardware isolation, unifying scattered multi-brand chip resources into standardized, low-cost and accessible AI token services, becoming the key hub connecting Ascend, Muxi, Moore Threads and all mainstream computing power resources.

Jingsuan Token Factory under Beijing Data Group realizes efficient pooling and scheduling of heterogeneous computing power. It raises the industry average resource utilization rate from 30%-40% to over 80%. Supported by the PD-separated intelligent scheduling architecture, it boosts the maximum computing throughput by 90%, improves large model inference latency by more than 70%, and cuts the single-token subscription cost for long-text scenarios by over 50%.

Tuoyuan, a domestic token optimization factory launched by Shikong Technology, is compatible with more than 10 domestic computing chips and over 20 mainstream AI models, with a daily token throughput of billions. It builds a full-link optimized AI inference system: matching the optimal computing power link through request portraits at the task adaptation layer, releasing full chip performance via exclusive operator fusion compilation at the operator library layer, and breaking cross-regional and multi-brand computing power isolation at the heterogeneous cluster scheduling layer.

SenseTime Galaxy Project plans to build 1 token operation center and 5 ten-thousand-card-level domestic intelligent computing clusters, integrating multi-vendor chips, computing clusters and model services into a unified standardized operation system. China Telecom Xingchen TokenHub integrates model access, computing scheduling, security audit and cost control into a one-stop service system, enabling convenient invocation via standard interfaces for government, enterprise and developer users.

All platforms share the core logic: completely shielding the underlying hardware differences of Ascend, Muxi, Moore Threads and other chips, and delivering unified, standardized token services to end developers and enterprises.

Unexhub: Pushing Tokenized Computing Power Trading to Industrial Extremes

Unexhub further iterates and upgrades the token trading mechanism, building a professional token aggregation and distribution computing power trading center. It optimizes the traditional computing power service model through three core mechanisms, bringing tangible benefits to both resource suppliers and end users.

Real-time settlement to eliminate cash flow pressure: Different from the long-cycle delayed settlement of traditional platforms, Unexhub’s independent financial clearing system enables suppliers to get paid immediately after service delivery. Revenue is settled in real time based on actual computing power output, with nearly zero capital occupation pressure.

Zero-commission incentive for early settlers: Against the industry’s universal 10%-30% service commission rate, Unexhub provides zero platform commission for all early-entry suppliers, rapidly expanding high-quality global AI resource supply through short-term revenue concessions.

Dynamic API bidding mechanism for market-oriented pricing: API service prices float freely based on real-time market supply and demand. Off-peak low prices help individual developers reduce R&D and experimental costs; peak demand brings reasonable profit premiums for suppliers, realizing transparent and fair market pricing.

Three-Tier Industrial Architecture Rewrites AI Inference Cost Rules

The large-scale adaptation of H3 forms a complete and clear industrial logic chain with open-source model dividends and token computing power trading:

First layer: Open-source models lower usage thresholds. H3 Base weights are permanently free to download and deploy, completely eliminating expensive model authorization fees. The ¥0.8/second 2K video generation cost has established a new industry cost benchmark.

Second layer: Full chip adaptation breaks hardware lock-in. The Day0 adaptation of 16 vendors supports free H3 deployment on all mainstream domestic computing platforms including Ascend, Muxi, Moore Threads, Hygon, Kunlunxin and Biren Technology, helping enterprises avoid single-vendor hardware dependence.

Third layer: Computing power trading realizes standardized low-cost delivery. Token factories and trading centers complete unified scheduling, metering and service output for multi-brand heterogeneous chips. Users only focus on token price stability and cost performance, without caring about underlying hardware brands.

The superposition of the three layers completely reconstructs the enterprise H3 application cost structure: free open-source models + optional full-brand hardware + market-bidding low-cost computing power. The synchronous adaptation of 16 chip vendors is not merely a technical achievement, but the first large-scale industrial verification of heterogeneous computing power unified scheduling and on-demand multi-modal AI services.

Industry Trends & Unexhub Core Advantages

Currently, only H3 Base is open-sourced, while H3 Context IR and H3 Regenerate-2K modules are not yet available for local deployment, and the complete workflow still relies on official APIs. Even so, the industry trend is irreversible: open-source models lower the "usage threshold", full chip adaptation lowers the "selection threshold", and token computing power trading lowers the "operation threshold". The integration of the three marks the official arrival of the productive era for multi-modal AI.

As a global token trading leader, Unexhub reshapes the international token circulation logic. It gathers full-spectrum AI 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, it delivers extreme cost-effective services.

For developers and enterprises: Token actual payment cost is as low as 30% of the market price, supporting low-cost batch calls and large-scale commercial deployment.

For AI resource 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.


16家芯片厂商同日适配H3:算力交易中心如何串联昇腾、沐曦、摩尔线程?

8月3日,MiniMax 旗舰级全模态生成模型 H3 正式开源,掀起全球 AI 软硬件生态全面升级浪潮。模型首发当日,16家海内外主流生态伙伴同步完成 Day0 首日全量适配,覆盖国内外头部芯片厂商、开发者社区、云推理平台及核心推理框架,生态落地速度刷新行业纪录。

本次同步适配阵容全面且硬核:包含华为昇腾、摩尔线程、沐曦、海光信息、昆仑芯、天数智芯、壁仞科技7大国产主力芯片厂商,AMD、英特尔两大国际芯片巨头,同时囊括 Hugging Face、魔搭 ModelScope、ComfyUI、RunningHub、fal 等主流开发社区与云推理平台,以及 vLLM-Omni、SGLang 核心推理框架。

这场全网同步适配绝非偶然,标志性的 Day0 无缝兼容能力,意味着全新 H3 模型上线即可在全品类主流平台稳定部署、直接运行,彻底告别以往新模型发布后,需数月等待硬件适配的行业常态。

各大厂商均交出亮眼适配成果:沐曦曦云 C 系列 GPU 率先完成 H3 首日适配;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,就打通 SGLang-MUSA 推理框架、MATE、muDNN 高性能算子库的全链路适配路径,适配效率行业顶尖。

国产高端算力芯片彻底追上全球前沿 AI 模型迭代节奏,从过去「滞后数月适配」实现跨越式升级,达成「发布即通跑」的行业新高度。

H3 能够实现16家厂商全域同步跟进,核心源于其「性能顶尖+商业价值突出」的双重核心优势。作为通用型全模态生成系统,H3 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、自带立体原声的超清视频输出。在 Artificial Analysis 权威音视频榜单中,H3 以1130的超高 Elo 评分稳居全球第一,综合实力超越 Gemini Omni Flash、万相2.7等国内外主流视频大模型。

相较于亮眼的技术性能,真正驱动全行业厂商全力首日适配的,是其颠覆性的商用性价比。H3 的2K超清视频生成成本低至0.8元/秒,仅为行业同类旗舰模型的三分之一。这份极致低价并非短期补贴造势,而是底层架构深度优化的长效成果:通过高压缩 Token 分词器大幅削减视频生成算力消耗,搭配异构训练优化、智能负载均衡、GPU 利用率提升等全维度系统迭代,从根源挤压冗余算力开销,构建稳固的行业成本优势。

开源免费的模型权限、极致压缩的推理成本、快速扩张的全域生态,三重优势叠加,让芯片厂商适配 H3 不再是被动的技术公益迭代,而是瞄准多模态推理蓝海市场、抢占行业新一轮增长红利的前瞻性战略布局。

行业核心痛点:硬件生态碎片化,开发门槛居高不下

16家厂商同步适配看似生态繁荣,却未能解决底层硬件碎片化难题。昇腾、沐曦、摩尔线程等各品牌芯片,均拥有独立的工具链、部署环境与参数调试体系。开发者想要高效部署、运行 H3 模型,需提前筛选硬件型号、搭建专属运行环境、反复调试适配参数,流程繁琐、技术门槛高、时间成本高,严重制约多模态 AI 的规模化落地。

算力交易中心:打通硬件孤岛,串联全品类算力资源

Token 算力交易中心的出现,完美破解异构算力孤岛痛点,将原本分散、独立的昇腾、沐曦、摩尔线程等多品牌芯片资源,整合为统一、标准化、普惠化的 Token AI 服务,成为串联全域算力生态的核心纽带。

北京数据集团旗下京算词元工厂,通过异构算力池化调度技术,将行业30%-40%的算力平均利用率大幅提升至80%以上。依托 PD 分离智能调度架构,算力吞吐峰值提升90%,大模型推理场景延迟降低70%以上,长文本场景单 Token 订购成本直降50%以上,实现效率与成本双重突破。

实控科技打造的国产自研词元优化工厂拓元,兼容10余款国产算力芯片、适配20余款主流 AI 模型,日均 Token 吞吐量达数十亿级别。平台搭建 AI 推理全链路优化体系:任务适配层通过请求画像精准匹配最优算力链路,算子库层依托硬件独占算子融合编译释放芯片极致性能,异构集群调度层彻底打通跨区域、多品牌算力壁垒,实现全域算力统一调度。

商汤发起的银河计划,规划建设1个词元运营中心+5个万卡级国产智算集群,将多厂商芯片、智算集群与模型服务整合为一体化标准运营体系。天翼云星辰词元Hub平台,聚合模型接入、算力调度、安全审计、成本管控全流程能力,政企客户与开发者可通过标准化接口一键调用,大幅降低使用门槛。

一众头部算力平台的核心逻辑高度统一:彻底屏蔽底层昇腾、沐曦、摩尔线程等硬件差异,对开发者和企业输出统一、标准化的 Token 服务,真正实现「底层算力无差别,上层服务标准化」。

Unexhub:极致迭代Token交易规则,重构全球算力流通生态

Unexhub 将词元算力交易逻辑推向行业极致,打造专业化Token聚合分发算力交易中心,依托三大核心机制优化行业痛点,为供需双方带来实打实的专属权益与价值升级。

即时结算,零资金占用压力:颠覆传统算力平台次月、隔月延迟结算模式,Unexhub 搭建独立金融清算体系,供应商算力产出即可实时回款,供多少、结多少,彻底解决中小算力供应商现金流紧张、资金积压难题。

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 推出重磅生态福利,所有早期入驻供应商全程享受0平台佣金,以短期平台收益让利,快速吸纳全球优质算力与模型资源,完善生态供给。

API市场化竞价,价格透明公正:打破平台单方面定价垄断,API 调用价格随市场供需动态浮动。闲时算力低价普惠,助力开发者低成本研发试错;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,杜绝隐形溢价。

三层产业架构闭环,彻底改写AI商用成本逻辑

H3 开源迭代、全域芯片适配、Token 算力交易,三者层层赋能、深度耦合,形成清晰完整的产业逻辑闭环,彻底重构企业多模态 AI 商用成本结构。

第一层:开源模型,清零授权成本。H3 Base 权重免费开源、可自由部署,彻底砍掉传统模型高额授权费用,0.8元/秒的2K超清视频生成成本,筑牢行业极致低价基准线。

第二层:全域适配,打破硬件锁定。16家厂商 Day0 同步适配,让企业可自由选择昇腾、沐曦、摩尔线程、海光、昆仑芯、壁仞科技等任意国产算力平台部署模型,彻底摆脱单一硬件供应商绑定,选型灵活无壁垒。

第三层:算力交易,普惠标准化落地。词元工厂、Unexhub 等算力交易平台,实现多品牌异构芯片的统一调度、统一计量、统一交付。开发者无需关注底层硬件型号,只需按需调用 Token 服务,聚焦成本与可用性。

三层架构叠加,打造出「免费模型+自由选型+竞价低价算力」的全新商用模式。16家芯片厂商同日适配 H3,绝非单纯的技术新闻,而是「异构算力统一调度、多模态AI按需普惠」产业趋势的首次大规模落地验证。

行业趋势与Unexhub核心权益

目前行业仍处于迭代初期:仅 H3 Base 模块开源,H3 Context IR、H3 Regenerate-2K 尚未开放本地部署,完整工作流仍依托官方 API。但行业发展大势已然清晰:开源模型降低 AI「使用门槛」,全域芯片适配降低「选型门槛」,Token 算力交易降低「运营门槛」,三者交汇融合,正式推动多模态 AI 迈入工业化生产力时代。

Unexhub 全面重塑全球 Token 流通格局,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 AI 资源,搭建海量全球竞价资源池。依托毫秒级智能路由技术,实时匹配全网最优低价算力,极致优化使用体验与成本。

用户专属权益:开发者、企业用户 Token 实付成本最低可至市场原价3折,高强度批量调用、大规模商用场景均可极致控本增效。

供应商专属权益:支持一键入驻平台,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。