16 家芯片厂商首日适配 H3:算力交易中心如何调度昇腾、沐曦、摩尔线程跑通多模态开源

On the exact day of release, a total of 16 mainstream ecological partners completed full Day0 adaptation, covering multiple dimensions including top chip manufacturers, open-source communities, cloud inference platforms, and inference frameworks.


16 Chip Vendors Adapt H3 Simultaneously: How Computing Power Trading Centers Connect Ascend, Muxi and Moore Threads

August 3 marked a milestone for the domestic multi-modal AI industry, as MiniMax officially open-sourced its flagship full-modal generation model H3. On the exact day of release, a total of 16 mainstream ecological partners completed full Day0 adaptation, covering multiple dimensions including top chip manufacturers, open-source communities, cloud inference platforms, and inference frameworks.

The participating ecological partners include 9 leading chip vendors: Huawei Ascend, Moore Threads, Muxi, Hygon Information, Kunlunxin, KUNLUN, Biren Technology, AMD, and Intel. They are joined by 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 large-scale synchronous Day0 adaptation is no coincidence. It means the brand-new H3 model can be deployed and run stably on mainstream hardware and platforms worldwide immediately upon release, ending the industry’s long-standing situation of "waiting months for model-chip adaptation".

Each chip vendor delivered impressive adaptation results. Muxi completed Day0 adaptation for MiniMax H3 with its self-developed Xiyun C-series GPU and MXMACA computing platform relying on independent hardware architecture capabilities. Moore Threads finished the full-stack adaptation path from the SGLang-MUSA inference framework to MATE and muDNN high-performance operator libraries in merely 3 hours. Hygon Information realized full-process adaptation, deployment and performance verification on its DCU platform, while Biren Technology also achieved seamless synchronous adaptation on the same day.

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

The industry-wide synchronous follow-up is rooted in H3’s dual advantages of top-tier performance and disruptive commercial value. As a leading universal full-modal generation system, H3 unifies contextual understanding of text, images, videos and audio, natively supporting up to 2K resolution, 15-second ultra-clear videos with built-in stereo audio. It ranks first globally with an Elo score of 1130 on the authoritative Artificial Analysis audio and video editing benchmark.

More importantly, H3 reshapes the commercial cost system of the industry with an ultra-low pricing of ¥0.8 per second for 2K video generation, only one-third of the cost of competing flagship models. This drastic cost reduction comes from systematic architectural optimizations: high-compression tokenizers cut down token consumption for video generation, while heterogeneous training, load balancing and GPU utilization upgrades substantially improve operational efficiency and squeeze marginal costs.

The combination of open-source accessibility, ultra-low cost and ecological expansion makes H3 adaptation a forward-looking strategic layout for chip vendors targeting the upcoming explosive growth of multi-modal inference markets, rather than a passive technical public welfare project.

Computing Power Trading Center: The Core Link That Integrates Heterogeneous Chip Resources

Despite the synchronous Day0 adaptation by 16 chip vendors, scattered hardware ecosystems once created obvious barriers for developers. Ascend, Muxi, and Moore Threads each have independent toolchains, deployment environments and parameter systems. Running H3 efficiently used to require developers to select hardware first, build exclusive environments, and debug parameters repeatedly, with high technical thresholds and adaptation costs.

The emerging token-based computing power trading centers completely solve the fragmentation pain point of heterogeneous hardware, unifying scattered chip resources into standardized and accessible AI services.

Major domestic trading platforms have built mature unified scheduling capabilities. The "Jingsuan Token Factory" supports efficient integration of diverse hardware relying on seven core capabilities including PD-separated intelligent scheduling, full-stack inference acceleration and heterogeneous computing power unified scheduling. China Unicom Xingluo Token Service Platform eliminates computing power isolation through network-cloud integration, realizing standardized and inclusive token services. SenseTime’s "Galaxy Project" integrates multi-vendor chips, intelligent computing clusters and model services into a unified operation system via self-built token operation centers and ten-thousand-card-level domestic computing clusters.

Practical industrial cases further verify the integration value. Silicon Flow aggregates heterogeneous computing resources including NVIDIA, AMD, Ascend, Muxi and Moore Threads through its self-developed SiliconLLM inference engine and computing orchestration system, converting fragmented hardware resources into standardized token supply. Tuoyuan Technology realizes unified scheduling and inference acceleration for domestic heterogeneous computing clusters with adaptive task optimization, exclusive operator fusion compilation and heterogeneous cluster scheduling technologies. It is currently compatible with more than 10 mainstream domestic chips, with a daily token throughput of billions.

All platforms share the core logic: underlying hardware differences between Ascend, Muxi, Moore Threads and other chips are completely shielded. Developers only need to access standardized token services without focusing on underlying hardware types, greatly lowering the threshold for multi-modal AI development.

Three-Tier Industrial Architecture: Open-Source Model + Chip Adaptation + Computing Power Transaction

The large-scale synchronous adaptation of 16 chip vendors forms a complete closed-loop industrial logic with H3 open-source iteration and token computing power trading, rewriting the commercial application system of multi-modal AI.

First layer: Open-source model releases technological dividends. H3 Base weights are permanently free to download and deploy, completely eliminating expensive model authorization fees. The ¥0.8/second 2K video generation cost anchors the new industry bottom price for high-quality multi-modal production.

Second layer: Full chip adaptation breaks hardware lock-in. The comprehensive Day0 adaptation of 16 mainstream vendors enables enterprises to deploy H3 freely on domestic high-performance computing platforms such as Ascend, Muxi, Moore Threads, Hygon and Kunlunxin, realizing flexible hardware selection and avoiding single-vendor dependence.

Third layer: Computing power transaction realizes standardized low-cost delivery. Token factories and trading centers complete unified scheduling, metering and service output for heterogeneous chips. Hardware differences are fully shielded, and developers only focus on token price, stability and availability, realizing on-demand use of AI computing power.

Unexhub: Ultimate Iteration of Global Token Trading Mechanism

Unexhub further upgrades the third-tier industrial logic, positioning itself as a global leading token aggregation and distribution computing power trading center. It optimizes the traditional computing power trading model with three core mechanisms, bringing tangible cost reduction and efficiency improvement benefits to both suppliers and developers.

Real-time settlement: Solving the long-standing cash flow pressure on suppliers. Different from the monthly or quarterly delayed settlement of traditional platforms, Unexhub’s independent financial clearing system supports real-time settlement after service delivery. Suppliers get paid immediately for valid computing power output with zero capital occupation pressure.

Zero-commission early incentive: Against the industry’s universal 10%-30% commission rate, Unexhub provides zero platform commission for all early settled suppliers. Short-term revenue concessions rapidly expand high-quality global AI resource supply and improve ecological maturity.

API dynamic bidding: Breaking unilateral platform pricing. API service prices float dynamically based on real-time market supply and demand. Off-peak low prices help developers reduce R&D costs, while peak demand brings reasonable profit premiums for suppliers, realizing market-oriented fair pricing.

Industrial Value & Future Trends

The superposition of the three-tier architecture completely reconstructs the enterprise H3 usage model: free open-source models + freely selectable heterogeneous chips + market-bidding low-cost computing power. The synchronous adaptation of 16 chip vendors is not merely a technical update, but the first large-scale industrial verification of heterogeneous computing power scheduling and on-demand multi-modal AI services.

Currently, only H3 Base is open-sourced, while H3 Context IR and H3 Regenerate-2K modules remain unavailable for local deployment, and the complete workflow still relies on official APIs. Even so, the industry trend is clear: 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 officially drives multi-modal AI from laboratory research to large-scale productive application.

Unexhub Exclusive Core Advantages

Unexhub reshapes the global token trading ecosystem, gathering 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. Equipped with millisecond-level intelligent routing and full-market bidding mechanism, it screens the lowest-cost and most stable AI resources in real time.

For developers and enterprises, actual token payment costs are as low as 30% of the market price, supporting low-cost batch calls and large-scale commercial production. For suppliers, one-click access is available with real-time settlement of 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 为国内多模态 AI 行业投下一枚重磅变革炸弹——新一代全模态生成模型 H3 正式开源。伴随模型首发,16家海内外主流生态伙伴同步完成 Day0 首日适配,覆盖头部芯片厂商、开源社区、云推理平台、推理框架四大核心赛道,刷新行业生态落地速度纪录。

本次同步适配阵容堪称豪华: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、AMD、英特尔9大主流芯片厂商;同时覆盖 Hugging Face、魔搭 ModelScope、ComfyUI、RunningHub、fal 等知名开发社区与云推理平台,以及 vLLM-Omni、SGLang 两大核心推理框架,实现软硬件生态全域打通。

这场全网同步适配绝非偶然,标志性的 Day0 首日适配,意味着 H3 模型一经发布,即可在全球主流硬件与平台上直接部署、稳定运行,彻底终结了以往新模型发布后,芯片适配滞后数月、落地困难的行业痛点。

各大厂商均交出亮眼适配成果:沐曦依托自研曦云 C 系列 GPU 硬件架构与 MXMACA 计算平台,率先完成 H3 首日适配;摩尔线程仅耗时3小时,就打通 SGLang-MUSA 推理框架、MATE、muDNN 高性能算子库的全栈适配链路;海光信息完成 DCU 平台全流程适配、部署与性能核验;壁仞科技同步完成无缝兼容适配,国产芯片生态全面跟上前沿模型迭代节奏。

国产算力彻底实现从「滞后适配、被动跟进」到「首日通跑、主动卡位」的质变,为多模态 AI 规模化落地筑牢硬件底座。

16家厂商全力跟进的核心原因,不止于 H3 顶尖的技术实力,更在于其颠覆性的商业性价比。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,原生输出最高2K分辨率、15秒时长、自带立体原声的超清视频,以1130的 Elo 评分登顶 Artificial Analysis 全球音视频榜单。

更具商业价值的是,H3 的2K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。这份极致低价并非补贴堆砌,而是全方位技术优化的结果:高压缩 Token 分词器大幅减少视频生成算力消耗,搭配异构训练架构、智能负载均衡、GPU 利用率优化,从底层架构层面持续压缩推理成本,打造长效降本优势。

开源免费、极致降本、生态爆发,三重优势叠加,让适配 H3 不再是芯片厂商的公益技术迭代,而是瞄准多模态推理蓝海市场的前瞻性战略布局,抢占行业新一轮增长红利。

算力交易中心:串联异构芯片生态的核心纽带

16家芯片厂商各自完成适配,看似生态繁荣,却存在显著的碎片化难题:昇腾、沐曦、摩尔线程均有独立的工具链、部署环境与参数体系。开发者想要高效运行 H3,必须先筛选硬件、搭建专属环境、反复调试参数,技术门槛高、适配成本高,严重制约模型规模化普及。

而Token算力交易中心的出现,完美破解异构算力孤岛问题,将分散、独立的各品牌芯片资源,整合为统一、标准化、低门槛的普惠 AI 服务,成为串联全行业硬件生态的核心纽带。

国内各大算力交易平台已形成成熟的统一调度能力:「京算词元工厂」依托PD分离智能调度、全栈推理加速、异构算力统一调度等七大核心能力,实现多品类算力高效融合;「联通星罗」词元服务平台以算网融合为核心,打通跨品牌算力壁垒,实现词元服务标准化、普惠化;商汤「银河计划」布局词元运营中心与五大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统一运营体系。

落地案例更能体现其实用价值:硅基流动依托自研 SiliconLLM 推理引擎与算力编排系统,聚合英伟达、AMD、昇腾、沐曦、摩尔线程等异构算力,统一转化为标准化 Token 供给;拓元科技通过任务自适应优化、硬件独占算子融合编译、异构集群调度等核心技术,实现国产异构算力池统一调度与推理加速,目前已兼容10余款主流国产芯片,日均 Token 吞吐量达数十亿级别。

所有平台的核心逻辑高度统一:彻底屏蔽昇腾、沐曦、摩尔线程等底层硬件差异,对开发者输出统一标准化的 Token 服务。用户无需关注底层硬件型号,只需按需调用服务、按量付费,彻底降低多模态 AI 开发门槛。

三层产业架构:开源模型+芯片适配+算力交易,重塑行业格局

16家芯片厂商同日适配、H3 开源迭代、Token 算力交易,三者深度耦合,形成一套完整的三层产业闭环,彻底改写多模态 AI 的商用落地逻辑。

第一层:开源模型,锚定极致成本底线。H3 Base 权重永久免费开源,彻底清零高额模型授权费用,0.8元/秒的2K超清视频生成成本,为行业高品质多模态内容生产锚定全新低价基准。

第二层:全域芯片适配,打破硬件锁定。16家主流厂商首日全覆盖适配,让企业可自由选择昇腾、沐曦、摩尔线程、海光、昆仑芯等国产算力平台部署 H3,摆脱单一硬件供应商绑定,实现硬件选型自由化。

第三层:算力交易,实现标准化普惠落地。词元工厂与算力交易中心完成多品牌异构芯片的统一调度、统一计量、统一交付,屏蔽硬件差异,让开发者只关注 Token 价格、稳定性与可用性,真正实现 AI 算力像水电一样按需取用。

Unexhub:极致升级全球Token交易规则

Unexhub 将第三层算力交易逻辑推向极致,定位为全球化Token聚合分发算力交易中心,区别于传统单一算力租赁平台,以三大核心机制优化行业生态,为供需双方带来实打实的专属权益。

即时结算,零资金压力: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,彻底解决中小供应商现金流占用难题。

早期零佣金,全力让利生态:行业通用佣金率普遍高达10%-30%,平台大幅挤压供需利润。Unexhub 重磅推出早期入驻福利,前期入驻供应商全程享受0平台佣金,以短期收益让利,快速扩容全球优质 AI 算力与模型资源。

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。闲时低价普惠开发者,降低研发试错成本;忙时需求上涨,供应商可获得合理溢价收益,定价权完全回归市场。

产业价值与未来趋势

三层架构层层赋能,彻底重构企业 H3 商用成本体系:免费开源模型 + 自由选型异构芯片 + 市场化竞价低价算力。16家芯片厂商的同日适配,不仅是一次技术生态更新,更是「异构算力统一调度、多模态AI按需服务」产业趋势的首次大规模落地验证。

目前产业仍处于迭代初期: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 走出实验室,迈入规模化生产力时代。

Unexhub 核心专属权益

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,搭建海量全球竞价池。依托毫秒级智能路由技术,实时筛选全网最优低价算力资源。

平台权益极具竞争力:开发者、企业用户 Token 实付成本最低可达市场原价3折,完美适配高强度批量调用、大规模商用场景,极致控本增效;算力与模型供应商支持一键入驻、实时结算分成、早期零佣金,高效灵活变现,是全球 AI 供需双方的最优合作枢纽。