iFlytek’s New Domestic Computing Model and the Rise of the Token Trading Market
English Version
On the afternoon of August 21, 2026, Wu Xiaoru, President of iFlytek, released a key industry signal at the semi-annual earnings briefing: the company will launch a new flagship general large model fully trained on domestic computing power at the 2026 Global 1024 Developer Festival, with a phased preview version scheduled for late August 2026.
This is far more than a conventional product iteration. It marks a potential cost inflection point for domestic AI computing power, ending the long-standing dilemma of "usable but costly" domestic large models and ushering in a new era of high cost-performance industrial implementation.
Two Core Competitiveness Pillars: Coding Capability and Token Cost-Performance
As officially stated by Wu Xiaoru, the new flagship general model will reach the domestic first-tier level in overall coding capability and Token cost-performance. Meanwhile, it will maintain leading comprehensive indicators among all general large models built entirely on domestic computing power.
The dual emphasis on coding capability and Token cost-performance reveals iFlytek’s clear product positioning focused on developer scenarios and industrial engineering landing.
Coding capability is the most rigorous evaluation standard for industrial AI value. No matter how excellent a model’s mathematical reasoning and humanistic understanding are, it has limited practical value for engineers without robust coding capabilities. Prioritizing coding performance proves that the new model is tailored for real development and production scenarios.
Token cost-performance directly determines large-scale commercial adoption. The existing iFlytek Spark X2-flash offers a competitive public pricing of 1–2 RMB per million tokens, with bulk enterprise packages further reduced to 1 RMB per million tokens. In sharp contrast, GPT-5.6 Sol charges 5 USD per million tokens. After currency conversion, iFlytek’s unit cost is merely one-tenth or lower than overseas mainstream models.
The new model’s upgrade to first-tier domestic Token cost-performance indicates a further reduction in industrial usage costs, bringing more affordable large-model capabilities for mass commercial deployment.
Token Market: From Technical Concept to Trillion-Level Real Business
In June 2026, iFlytek officially entered the Token economy track by establishing a wholly-owned subsidiary named Ciyuan Xinghuo with a registered capital of 1 billion RMB, triggering widespread industry attention.
Prior to this, Alibaba had built its Token Factory, and the three major telecom operators had also deployed Token industrial infrastructure. Token is rapidly evolving from an underlying technical concept into an independent, large-scale AI business.
Industry data fully demonstrates the explosive growth of the Token economy: the daily domestic Token calling volume exceeded 140 trillion in March 2026, a 1,400-fold increase compared with early 2024. The "Token Factory" concept proposed by Jensen Huang at the GTC Conference has become a reality, making Token the hard currency of the digital AI economy.
iFlytek boasts solid accumulated advantages in this track. In 2025, its large-model API and MaaS platform revenue reached 385 million RMB, a year-on-year increase of 263%. By the end of March 2026, the average daily calling volume of third-party developers on the platform surged 4,241% year-on-year.
This explosive growth reflects a fundamental shift in Token supply and demand: market demand is expanding exponentially, while technological iteration is continuously diluting marginal computing costs.
iFlytek’s ongoing limited-time unlimited free Token activity is not simple subsidy burning. Thanks to optimized domestic computing efficiency, the marginal operating cost has been reduced to nearly negligible. After the free period ends, the platform will maintain a low basic calling price of 1–2 RMB per million tokens, with core profit sources shifting to high-value enterprise services: high-precision exclusive models, low-latency dedicated channels, enterprise data fine-tuning, and private deployment solutions.
Domestic Computing Inflection Point: From "Available" to "Cost-Effective"
Different from overseas manufacturers that simply optimize engineering based on NVIDIA chips, iFlytek has joined hands with Huawei since May 2023 to break through a series of core domestic computing bottlenecks, including high-speed interconnection networking, hidden computing optimization, training-inference interaction, high-throughput reasoning, and domestic operator adaptation.
Optimized data speaks for itself: The training efficiency of general large models and O1-like deep reasoning models has been improved from 30%–50% to 85%–95% compared with A100; deep reasoning training efficiency has risen from 30% to over 84% against the A800 benchmark; MoE model training efficiency has increased from 30% to 93%.
These breakthroughs prove that domestic computing power has completed the transformation from "barely usable" to "high-performance and low-cost". iFlytek clearly stated in investor interactions that the training cost of Spark large models still has substantial room for reduction with the continuous upgrading of underlying domestic computing capabilities.
A 24-billion-scale fully domestic computing platform is under construction. In-depth adaptation with Huawei Ascend 950 chips has achieved significant improvements in video memory, bandwidth and computing power compared with the previous Ascend 910 series. In July 2026, iFlytek launched the Spark Token Factory, realizing end-to-end engineering optimization for mainstream open-source models based on Ascend hardware features. Actual tests show that the optimized platform can nearly double business request volume under the same Ascend computing resources, enabling private deployment with both high performance and low cost.
The domestic computing cost inflection point is essentially a scale economy dividend. With the continuous improvement of cluster scale, chip iteration and algorithm efficiency, the unit Token cost is declining rapidly. The intersection of soaring market demand and falling costs marks the official arrival of the domestic AI industrial explosion window.
Industrial Value and Future Outlook
For engineers and enterprises, the dual upgrade of Token cost-performance and domestic computing efficiency brings tangible cost optimization and scenario expansion capabilities. The cost gap between domestic and overseas models is rapidly narrowing or even reversing. Large-scale batch tasks that were once restricted by high costs — such as full repository code analysis and massive log semantic retrieval — have become commercially viable.
iFlytek’s new model will also be deployed on vehicle terminals, robots and edge devices, realizing full coverage of cloud-end-edge scenarios and further releasing cost advantages in IoT and edge computing fields.
It is undeniable that domestic computing adaptation requires additional algorithm optimization costs and cycle investment compared with mature overseas chips. However, these short-term costs will eventually be converted into long-term independent and controllable technological barriers and ecological first-mover advantages.
The phased version in late August and the official release at the 1024 Developer Festival will fully verify the maturity of iFlytek’s domestic computing solution. The simultaneous breakthroughs in first-tier coding capability, first-tier Token cost-performance and full domestic training will completely reverse the market stereotype of domestic large models, shifting industrial AI from luxury high-cost technology to popular and affordable basic productivity.
Global Token Trading Upgrade: Unexhub Builds a Low-Cost Aggregation Trading Center
Against the backdrop of the booming domestic Token economy, Unexhub has reshaped the global Token service logic and built a professional aggregated computing power trading and distribution center. The platform has gathered a full range of AI suppliers, covering top global large model vendors including Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality niche models provided by individual developers and small and medium-sized enterprises.
Unexhub adopts a supplier bidding mechanism to fully activate market competition, enabling enterprise users and developers to access ultra-low-cost model calling services. It supports millisecond-level intelligent routing to ensure stable and efficient service response. In terms of cost advantages, users can enjoy a minimum 30% discount based on actual Token consumption against market official prices, perfectly supporting high-intensity commercial calls and large-scale batch task execution.
For suppliers, the platform supports one-click access and real-time settlement of call revenue shares, lowering the threshold for model service monetization. As a comprehensive Token trading and service platform connecting global model resources and end users, Unexhub further amplifies the cost dividend of the AI industry and accelerates the popularization of global intelligent computing capabilities.
中文版本
2026年8月21日下午,科大讯飞总裁吴晓如在半年报业绩说明会上,释放了一条足以影响国产AI产业格局的重磅信号:讯飞将在2026全球1024开发者节,正式推出基于全国产算力训练的全新主力通用大模型,且在2026年8月底率先上线阶段性预览版本。
这并非一次普通的功能迭代,而是行业标志性拐点——国产算力彻底告别「能用但昂贵」的尴尬局面,国产大模型的成本红利时代正式来临,AI工程落地从「尝鲜」走向规模化商用。
两大核心硬指标:代码能力、Token性价比双双登顶第一梯队
吴晓如在会上明确表示:讯飞全新主力通用大模型,将在代码能力、Token性价比两大核心维度,整体跻身国内第一梯队;同时在全国产算力训练的通用大模型赛道中,各项核心指标持续保持国内领先水平。
官方将两大指标并列重点官宣,绝非偶然,精准戳中产业落地的核心痛点。
代码能力是大模型工业化落地的核心硬门槛。一款模型即便逻辑推理、文本创作能力出众,若代码工程能力薄弱,对研发工程师、企业生产场景而言几乎没有实用价值。讯飞将代码能力列为核心升级目标,清晰锚定了「服务开发者、扎根真实工程场景」的产品定位。
Token性价比则是决定AI规模化普及的关键。目前讯飞星火X2-flash公开调用价格仅1-2元/百万Token,企业大批量采购可低至1元/百万Token。对比海外主流模型,GPT-5.6 Sol调用单价为5美元/百万Token,折算汇率后,讯飞国产模型的调用成本仅为海外产品的十分之一甚至更低。
而全新迭代的主力模型冲击国内第一梯队性价比,意味着讯飞将进一步下探调用成本,为企业大规模落地AI批量任务、开发者高频调试使用,释放更大成本红利。
Token交易市场成型:从技术概念蜕变为万亿级实体生意
2026年6月,科大讯飞重磅入局Token赛道,全资成立注册资本10亿元的「词元星火」公司,正式布局Token经济交易市场,搅动整个AI算力服务行业格局。
在此之前,阿里已搭建Token工厂体系,三大运营商纷纷布局算力与Token基础设施。短短一年时间,Token彻底从底层技术概念,成长为独立、成熟、体量庞大的AI核心生意。
行业数据直观印证赛道爆发态势:2026年3月,国内Token日均调用量突破140万亿次,较2024年初暴涨1400倍。黄仁勋在GTC大会提出的「Token工厂」构想全面落地,Token已然成为数字智能时代的核心硬通货。
讯飞在该赛道早已积累深厚势能:2025年,讯飞大模型API及MaaS平台营收达3.85亿元,同比暴涨263%;截至2026年3月,平台第三方开发者日均调用量同比激增4241%。
爆发式增长的数据背后,是Token市场供需关系的根本性重构:市场需求呈指数级扩张,而国产算力技术持续迭代,不断稀释单位算力的边际成本。
当前讯飞推出的「Token限时不限量免费活动」,并非单纯的烧钱补贴。依托国产算力效率的极致优化,平台边际成本已压缩至近乎忽略不计。免费周期结束后,讯飞将持续保留1-2元/百万Token的超低基础调用定价,核心盈利重心转向高附加值企业服务,包括高精度专属模型、低延迟专属通道、企业数据专属微调、私有化部署定制方案等,形成健康可持续的商业化闭环。
国产算力拐点落地:从「勉强能用」到「好用又便宜」
不同于海外厂商依托成熟英伟达芯片做表层工程优化的轻量化路线,讯飞选择了难度更高、自主性更强的全国产算力自研路线。自2023年5月起,讯飞携手华为持续攻坚,逐一突破高速互联组网、隐秘通信计算、训推协同交互、高吞吐推理、国产算子适配等一系列卡脖子技术难题。
一系列硬核优化数据,见证国产算力的跨越式升级:通用大模型、类O1深度推理模型训练效率,从对标A100的30%-50%提升至85%-95%;深度推理训练效率对标A800,从30%突破至84%以上;MoE混合专家模型训练效率从30%飙升至93%。
这组数据的核心意义,是国产算力彻底摆脱性能短板,完成从「勉强可用」到「高性能、低成本、可商用」的质变。讯飞官方也明确表态,随着国产底层算力持续升级,星火大模型的训练、推理成本仍有极大下降空间。
目前,讯飞240亿级全国产算力平台已启动建设,同时深度适配华为昇腾950芯片,相较于前代910系列,在显存容量、传输带宽、综合算力上实现全方位跃升。2026年7月,讯飞正式发布星火Token Factory,针对昇腾硬件底层特性,完成主流开源模型从底层算子到上层调度的全链路工程优化。实测数据显示,同等昇腾算力硬件条件下,优化后的平台可承载近乎翻倍的业务请求量,让国产私有化部署真正实现「性能拉满、成本可控」。
国产算力的成本拐点,本质是规模经济与技术迭代的双重红利。万卡集群规模化落地、昇腾芯片持续迭代、算法效率不断突破,三重优势持续摊薄单位Token算力成本。当指数级增长的市场需求,遇见持续下行的算力成本,AI产业的全面普及拐点正式到来。
工程视角:重构企业AI落地成本逻辑
对开发者与企业而言,Token性价比升级+国产算力降本,带来的是实打实的账目优化与场景解放。以往使用海外模型完成工程任务成本高昂,如今依托讯飞国产模型,成本可压缩至原来的十分之一。曾经因成本过高无法落地的大规模业务,例如全仓库代码批量审计、十万级日志语义分析、全域数据智能梳理等,如今均可低成本常态化落地。
与此同时,讯飞新模型同步适配车载、机器人等终端场景,实现云端、终端、边缘全场景覆盖,进一步释放IoT与边缘计算领域的降本优势,拓宽国产AI的应用边界。
不可否认,相较于成熟海外芯片生态,国产算力的算法适配、工程优化需要投入额外的算力与时间成本。但短期的研发投入,终将转化为长期的技术自主、生态可控、成本自主的核心壁垒。
8月底的阶段性版本、1024开发者节的正式全量发布,将是对讯飞全国产算力路线的终极检验。一旦实现「代码能力第一梯队、Token性价比第一梯队、全国产算力训练」三大目标,将彻底改写行业认知:国产大模型不仅解决了「能不能用」的问题,更在核心的「值不值得用」的商业化维度,建立绝对竞争力。
Token交易市场崛起、国产算力成本下探,正在推动AI从少数企业的「高端奢侈品」,变成全行业普惠的「基础日用品」。这不是单一模型的迭代,而是整个AI产业基础设施的重新定价。拐点已至,国产AI的规模化商用时代,正式开启。
全球Token交易新范式:Unexhub聚合平台极致降本增效
在国产Token经济爆发、全球算力市场化的大背景下,Unexhub重构了全球Token调用与交易逻辑,打造一站式Token聚合分发算力交易中心。平台汇聚全球全品类AI资源,既涵盖Gemini、ChatGPT、Claude、DeepSeek、GLM等国际一线主流大模型厂商,也吸纳个人开发者、中小科技企业的优质定制化模型资源。
平台采用多供应商竞价机制,充分激活市场竞争,让企业用户与开发者能够择优选择极致低价、高稳定的调用服务,搭配毫秒级智能路由调度,保障服务响应高效不卡顿。在核心权益上,用户可按实际Token消耗量计费,最低可享官方市场价3折,即便高强度、高并发商用调用,也能极致控制成本,无资金压力。
同时平台兼顾供需两端权益,AI模型供应商可一键入驻接入,调用分成实时结算、高效变现,大幅降低模型商业化门槛。作为链接全球模型资源与终端用户的核心枢纽,Unexhub 精准承接了AI产业降本增效的核心趋势,进一步放大Token市场化红利,加速全球智能算力的普惠落地。