从讯飞四 Agent 到微软 CARE,医疗 agent 应用生态正在重写诊室流水线

In July 2026, at the China Hospital Information Network Conference in Beijing, iFLYTEK Healthcare introduced AI Clinical Assistant 2.0.

From iFLYTEK’s Four-Agent System to Microsoft CARE: Medical AI Agents Are Rewriting the Clinical Workflow

Bilingual Article|English + Chinese
The complete English version appears first.
中文完整版位于文章后半部分,请继续向下阅读。

In July 2026, at the China Hospital Information Network Conference in Beijing, iFLYTEK Healthcare introduced AI Clinical Assistant 2.0.

The system is built around four specialized AI agents, each responsible for a different stage of the clinical workflow.

The first listens to conversations between doctors and patients, then generates outpatient records, ward-round notes, surgical records, and discharge summaries.

The second reviews those records for missing information, logical inconsistencies, and potential quality risks.

The third combines medical history, laboratory findings, and examination results to identify possible diagnostic directions and high-risk conditions.

The fourth matches clinical guidelines and medical research, creating a traceable evidence base for the model’s recommendations.

Four tasks.

Four agents.

One coordinated clinical workflow.

This is a meaningful departure from the previous generation of medical AI.

In the past, hospitals often deployed isolated systems: one application generated medical records, another warned about medication risks, and a third searched clinical guidelines. Doctors had to move repeatedly between screens, copy information across platforms, and manually connect the results.

iFLYTEK is attempting to bring these capabilities into one continuous process, allowing several agents to collaborate around the same patient instead of operating as disconnected software modules.

According to the released deployment data, the outpatient system has already generated more than 100,000 medical records, with an adoption rate of approximately 91%. Each physician reportedly saves an average of 2.8 hours per day on medical-record writing.

In inpatient settings, the efficiency of producing ward-round records and discharge summaries has reportedly improved by 98% and 92%, respectively.

These are not merely laboratory demonstrations.

They reflect how medical agents are beginning to produce measurable results inside major hospitals.


Medical AI Is Moving From Isolated Tools to Coordinated Agent Teams

iFLYTEK is not the only company redesigning the clinical workflow.

In June, Tencent Health introduced a full-stack AI healthcare solution built around a four-layer architecture, covering medical services, insurance, assisted diagnosis, and 14 healthcare-specific agent capabilities.

For hospitals, its open platform provides six major AI functions:

  • Intelligent patient guidance;

  • Medical question answering;

  • AI-assisted pre-consultation;

  • Medical-report interpretation;

  • Medication planning;

  • Patient follow-up and recall.

For pharmaceutical companies, Tencent’s NGES 2026 solution combines the Chonglou foundation model with MedAgent Suite.

The platform provides approximately 11 standardized skill modules that can be freely combined according to business requirements.

Instead of developing every medical workflow from scratch, pharmaceutical companies can assemble reusable capabilities and shorten customization cycles from several days to a few hours.

This is an important signal.

Medical AI is no longer being packaged as one general-purpose model that attempts to do everything.

It is increasingly becoming an ecosystem of specialized agents, tools, and skills that can be combined around a specific clinical process.


Microsoft CARE: Turning Medical Reasoning Into an Auditable Evidence Chain

Microsoft is working at a more fundamental level.

Microsoft Research Asia and Yale University proposed CARE, a framework accepted by ICLR 2026.

CARE attempts to solve one of the most difficult problems in medical AI: the black-box nature of image-based diagnosis.

Traditional medical imaging models may produce a conclusion without clearly explaining how that conclusion was reached.

CARE restructures the reasoning process into three stages:

  1. Observation
    Identify relevant anatomical structures or suspicious lesions.

  2. Localization
    Use specialized segmentation tools to locate the important region precisely within the image.

  3. Diagnosis
    Combine the verified regions with medical evidence to produce the final conclusion.

Instead of asking users to trust an unexplained answer, CARE creates a complete reasoning path that can be inspected and audited.

The model does not simply say:

“This image may indicate a particular disease.”

It can show:

  • Which anatomical structure was examined;

  • Where the suspected abnormality was located;

  • Which region influenced the conclusion;

  • What evidence supported the final diagnosis.

This is crucial in healthcare.

A clinical AI system cannot rely only on accuracy scores. It must also allow physicians, hospitals, and regulators to understand how a recommendation was generated.

CARE transforms medical-image reasoning from a black-box prediction into a traceable evidence workflow.


Dragon Copilot Is Also Evolving Into a Unified Clinical Assistant

At HIMSS 2026, Microsoft announced that Dragon Copilot had evolved beyond a standalone speech and documentation tool.

It is becoming a unified AI clinical assistant connected with Microsoft 365 Copilot.

The system can support a broader range of work, including:

  • Clinical documentation;

  • Medical conversation capture;

  • Information retrieval;

  • Workflow assistance;

  • Follow-up tasks;

  • Collaboration across Microsoft applications.

Microsoft also introduced Copilot Health in March, allowing users to import electronic medical records from more than 50,000 hospitals and healthcare institutions in the United States.

The direction is increasingly clear.

Microsoft is not treating medical AI as one isolated diagnostic model.

It is building an agent layer that connects clinical data, documentation, medical evidence, and workplace applications.


The Industry Focus Is Shifting From Models to Agents

The central question in medical AI is changing.

For several years, the industry focused on foundation models:

  • Which model understands medicine better?

  • Which one scores higher on medical exams?

  • Which one can process longer medical records?

  • Which one has more parameters?

Now, the discussion is shifting toward agents:

  • Which agent listens to the consultation?

  • Which one writes the medical record?

  • Which one checks quality?

  • Which one searches clinical evidence?

  • Which one follows up with the patient?

  • Which one connects to the hospital information system?

At CHIMA 2026, IDC summarized the problem directly:

The supply of AI technology is beginning to exceed hospitals’ capacity to absorb it.

The largest gap is no longer simply model capability.

It is the ability of hospital systems to connect with, manage, and govern AI agents.

Many hospitals still operate fragmented information systems:

  • Electronic medical records;

  • Laboratory systems;

  • Imaging systems;

  • Pharmacy platforms;

  • Insurance systems;

  • Patient-management platforms;

  • Administrative databases.

Even if an AI agent performs well in isolation, it cannot create real value if it cannot access the correct data, return results to the correct workflow, or operate under clear permission controls.

The new challenge is not only building a smarter model.

It is building an infrastructure capable of supporting an entire agent ecosystem.


The Hidden Question: Who Supplies the Compute and Tokens?

Once medical agents begin operating at scale, another question becomes unavoidable:

Who provides the computing power and Tokens required to keep the system running?

Consider the scale of the emerging workload.

iFLYTEK’s outpatient system generates more than 100,000 medical records.

Tencent’s 14 medical-agent capabilities may operate across different parts of the same care process.

Microsoft CARE invokes several expert models and tools across observation, localization, and diagnosis.

A single consultation may trigger:

  • Speech recognition;

  • Medical-record generation;

  • Quality review;

  • Risk analysis;

  • Guideline retrieval;

  • Evidence verification;

  • Follow-up planning.

Each stage consumes model calls, context, tool operations, and computing resources.

As the number of agents increases, Token consumption does not grow in a simple linear pattern.

Agents may repeatedly read the same records, exchange intermediate results, invoke specialist models, verify outputs, and retry failed steps.

The total cost can rise rapidly.

Large hospitals may be able to negotiate dedicated infrastructure and long-term cloud contracts.

Smaller hospitals, community clinics, independent medical developers, and regional healthcare providers face a much more difficult equation.

They need access to advanced models, but may not have the budget or technical resources to maintain separate connections to multiple providers.

This is where the Token supply and settlement layer becomes part of the medical AI infrastructure.


Unexhub: Rebuilding the Global Token Distribution Layer

Unexhub is attempting to address this problem by restructuring how Tokens and AI computing resources are aggregated, distributed, priced, and settled.

At its core, Unexhub is building a computing-power trading center focused on Token aggregation and distribution.

The platform connects fragmented supply from model and computing providers with demand from enterprises, developers, and AI applications.

For medical-agent ecosystems, this role is increasingly relevant.

A hospital may need:

  • One model for medical-record generation;

  • Another for long-context analysis;

  • A multimodal model for medical images;

  • A low-cost model for classification;

  • A premium model for high-risk review;

  • Backup capacity when the primary service is unavailable.

Maintaining separate contracts, accounts, interfaces, and billing structures for every provider is expensive and operationally complex.

A unified distribution platform can simplify the process.


Immediate Settlement Reduces Pressure on Smaller Suppliers

For small and medium-sized AI suppliers, one of the most difficult challenges is cash flow.

They provide computing capacity or model services immediately, but payments may not arrive until weeks or months later.

During that period, suppliers still need to pay for:

  • GPUs;

  • Electricity;

  • Bandwidth;

  • Cloud infrastructure;

  • Operations;

  • Engineering teams.

A long settlement cycle can place enormous pressure on smaller participants, even when demand is growing.

Unexhub’s settlement and clearing system is designed around near-real-time revenue distribution:

When computing power is supplied and consumed, the corresponding revenue can be settled more quickly.

This can help suppliers shorten the gap between providing capacity and receiving payment.

For smaller model providers and independent computing suppliers, faster settlement is not merely a convenience.

It can reduce working-capital pressure and allow them to reinvest in additional capacity.

Unexhub is also preparing an incentive system for early suppliers.

Eligible early participants may receive access to platform rates as low as 0% commission during the promotional stage, helping them enter the distribution network without surrendering a large share of initial revenue.

The value of this benefit is not simply “free access.”

It gives early suppliers:

  • A lower channel-acquisition cost;

  • Faster access to enterprise demand;

  • Shorter settlement cycles;

  • More room to test pricing;

  • A clearer path to recover infrastructure investment.


API Bidding Brings Market Pricing to Tokens

Another important component is Unexhub’s planned API bidding model.

Traditional API pricing is usually fixed by the model provider.

Users pay the published rate regardless of whether capacity is abundant or underused.

Unexhub introduces a more market-oriented pricing mechanism, in which API prices can change according to supply and demand.

During off-peak periods, developers may access idle capacity at lower prices.

During periods of high demand, suppliers may earn more by providing scarce resources.

This creates benefits on both sides.

For developers and enterprises:

  • Lower off-peak API prices;

  • More flexible routing options;

  • Access to a wider range of models;

  • Reduced dependence on one provider;

  • Better control over high-volume Token spending.

For suppliers:

  • Higher utilization of idle computing resources;

  • More opportunities to reach global demand;

  • Dynamic pricing during peak periods;

  • Faster settlement of usage revenue;

  • Lower distribution costs through one access channel.

This is not merely a traditional discount program.

It transfers part of the Token-pricing process from a fixed vendor list price to a market-based allocation mechanism.


From Fixed API Prices to a Token Marketplace

The development of medical agents makes this shift more urgent.

A hospital does not always need the most expensive model.

Different stages of the clinical workflow have different requirements:

Medical taskSuitable model type
Speech transcriptionFast, low-cost model
Medical-record draftingBalanced language model
Record-quality inspectionSpecialized review model
Medical-image analysisMultimodal medical model
High-risk diagnostic reviewPremium reasoning model
Guideline retrievalSearch and evidence agent
Patient follow-upLow-cost conversational model

If every stage is routed to the most expensive flagship model, costs can become unsustainable.

A more rational system dynamically selects resources according to:

  • Clinical risk;

  • Task complexity;

  • Required latency;

  • Data type;

  • Budget;

  • Current supply;

  • Service availability.

This is the logic of a Token marketplace.

The platform does not simply offer one cheaper API.

It attempts to match each task with the most appropriate model and computing resource.


Unexhub’s Model Aggregation and Millisecond Routing

Unexhub is designed to aggregate AI suppliers at different scales.

The platform can connect access channels for mainstream international and domestic models, including:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Other large and specialized AI models.

It can also onboard high-quality models and computing resources provided by individuals, startups, and small or medium-sized companies.

This diverse supply base creates a wider range of options for developers and enterprises.

Through millisecond-level routing, requests can be directed according to factors such as:

  • Model capability;

  • Price;

  • Latency;

  • Availability;

  • Region;

  • Task type;

  • Service quality.

For medical applications, routing is especially important.

A low-risk administrative task should not consume the same expensive resources as a high-risk diagnostic review.

A unified routing layer can help reserve premium capacity for the work that truly requires it.


Pay by Actual Token Usage

Unexhub supports usage-based payment, allowing developers to pay according to actual Token consumption.

This can reduce the need to pre-purchase large fixed packages from several providers.

For independent developers, startups, and smaller medical institutions, the benefit is greater flexibility:

  • Start with a smaller budget;

  • Test several models;

  • Scale usage according to demand;

  • Avoid idle prepaid capacity;

  • Switch models without rebuilding the entire application.

Under selected supply conditions and promotional resource plans, effective API prices may be as low as approximately 30% of standard market rates.

For high-volume medical-agent workflows, this can materially change the economics.

The benefit is not simply saving a few cents on one call.

It may allow the same budget to support:

  • More consultations;

  • More generated records;

  • More evidence retrieval;

  • More independent verification;

  • Longer agent workflows;

  • Wider deployment across smaller hospitals.

Of course, healthcare organizations should evaluate more than price.

Any Token or API platform serving medical scenarios must also be assessed for:

  • Data security;

  • Privacy protection;

  • Model reliability;

  • Regional compliance;

  • Auditability;

  • Service continuity;

  • Failure recovery;

  • Access controls.

Low cost becomes valuable only when it is combined with secure and stable infrastructure.


Two Industrial Shifts Are Converging

When the medical-agent and Token-market developments are viewed together, the direction becomes clear.

iFLYTEK’s four-agent system, Microsoft CARE, Tencent’s full-stack solution, and Dragon Copilot are turning the clinical process into a coordinated AI workflow.

Medical-record generation, quality inspection, diagnostic support, evidence retrieval, and patient follow-up are increasingly handled by specialized agents.

Doctors can spend less time on repetitive documentation and more time on clinical judgment and patient communication.

But this automated workflow can operate only if computing resources remain available.

The underlying system needs:

  • Stable model access;

  • Sufficient Token capacity;

  • Dynamic routing;

  • Transparent pricing;

  • Fast settlement;

  • Backup providers;

  • Reliable infrastructure.

Unexhub is working on this underlying layer.

Medical-agent companies are redefining what the clinical workflow should do.

Unexhub is attempting to redefine how the computing power required by that workflow is supplied and settled.

One side is redesigning the work.

The other is redesigning the infrastructure that allows the work to continue.

The real industrial transformation appears where these two layers meet.


The Challenges Have Not Disappeared

The direction is clear, but the transition will not be simple.

Hospitals still need to answer difficult questions:

  • Can existing systems support multiple agents?

  • Can data move safely across departments?

  • Are diagnostic recommendations sufficiently explainable?

  • Can every clinical conclusion be traced to evidence?

  • Who is responsible when an AI recommendation is wrong?

  • Can smaller institutions afford the required Token volume?

  • Can suppliers guarantee stable capacity during peak periods?

Medical AI cannot be deployed like an ordinary office chatbot.

It operates in a high-risk environment where errors may affect health and safety.

Therefore, the future medical-agent ecosystem will require several layers of governance:

  • Human clinical oversight;

  • Evidence traceability;

  • Permission controls;

  • Independent validation;

  • Model monitoring;

  • Secure data handling;

  • Reliable Token and computing supply.

The technology is moving from isolated intelligence toward coordinated workflows.

Governance and infrastructure must move with it.


The Bottom Line

The clinical workflow is being rewritten.

iFLYTEK is using four specialized agents to connect documentation, quality control, diagnostic support, and evidence retrieval.

Tencent is building modular medical-agent capabilities across hospitals, insurance, and pharmaceutical workflows.

Microsoft CARE is turning medical-image reasoning into an auditable evidence chain.

Dragon Copilot is evolving into a unified clinical assistant connected to a broader productivity ecosystem.

These systems answer one question:

What should medical agents do?

Unexhub is focused on another:

How can the required models, computing resources, Tokens, prices, and payments be organized efficiently?

Its Token aggregation and distribution model aims to connect global AI supply with enterprise and developer demand.

Its settlement system can help suppliers receive usage revenue more quickly.

Its planned bidding mechanism can give developers access to lower off-peak prices while allowing suppliers to benefit from peak demand.

Its unified model access and millisecond routing can help different tasks reach different resources.

Medical agents are rebuilding the clinical workflow.

Unexhub is rebuilding the Token distribution logic underneath it.

One is changing how healthcare work is performed.

The other is changing how the computing power behind that work is delivered.

The future of medical AI will not be shaped by models alone.

It will be shaped by the agents that coordinate care—and the infrastructure that keeps those agents running.


中文版

从讯飞四智能体到微软 CARE:医疗智能体生态正在重写诊室流水线

2026 年 7 月,北京,CHINC 大会。

科大讯飞医疗正式发布 AI 诊疗助理 2.0。

这套系统由四类专业智能体组成,分别承担诊疗流程中的不同任务。

第一类智能体负责听取医患对话,自动生成门诊病历、查房记录、手术记录和出院小结。

第二类智能体负责审查病历,识别缺失项目、逻辑冲突以及潜在的质量风险。

第三类智能体综合患者病史、检验结果和检查信息,提示可能的诊断方向与高风险情形。

第四类智能体匹配临床诊疗指南与医学研究,为模型输出建立可追溯的医学证据链。

四件事。

四个智能体。

一条连续运转的诊疗流水线。

这与过去一代医疗 AI 有着明显区别。

过去,医院中的 AI 往往以相互独立的软件形式存在:一套系统生成病历,另一套系统提示用药风险,还有一套系统检索诊疗指南。

医生不得不在多个页面之间反复切换,将不同系统的结果重新整理到一起。

科大讯飞正在尝试把这些能力整合进同一条诊疗流程,让多个智能体围绕同一位患者持续协作,而不是彼此割裂地运行。

根据已披露的落地数据,门诊场景已生成超过 10 万份病历,引用率约为 91%,每位医生每天平均可节省约 2.8 小时的病历书写时间。

住院场景下,查房记录和出院小结的撰写效率分别提升约 98% 和 92%。

这已经不是实验室中的演示。

医疗智能体开始在大型医院中产生可以被量化的真实产出。


医疗 AI 正从单点工具走向智能体团队

重写诊疗流水线的并不只有科大讯飞。

6 月,腾讯健康推出 AI 全栈解决方案,以四层架构覆盖就医、医保和辅助诊断等场景,并提供 14 项医疗专属智能体能力。

面向医院端,其开放平台提供六项主要 AI 能力:

  • 智能导诊;

  • 医疗智能问答;

  • 智能预问诊;

  • 医疗报告解读;

  • 用药规划;

  • 患者召回与随访。

面向药企端,腾讯 NGES 2026 以“重楼”基座模型与 MedAgent Suite 为核心,提供约 11 个可自由组合的标准 Skill 模块。

药企不必再为每一个医疗流程从零开发系统,而可以根据业务需求,快速组合已有能力,将部分定制开发周期从数天缩短至数小时。

这释放了一个非常明确的信号。

医疗 AI 不再只是一个试图包办所有工作的通用大模型。

它正在变成由专业智能体、工具和技能模块共同构成的生态系统,并围绕具体医疗流程进行组合。


微软 CARE:把医学推理变成可审计的证据链

微软的动作更加深入底层。

微软亚洲研究院与耶鲁大学提出 CARE 框架,相关论文被 ICLR 2026 收录。

CARE 试图解决医疗 AI 中最棘手的问题之一:医学影像推理的黑箱化。

传统医学影像模型可能直接给出诊断结论,却无法清楚说明这个结论究竟如何产生。

CARE 将推理过程拆解为三个阶段:

  1. 观察
    识别相关解剖结构或疑似异常病灶。

  2. 定位
    借助专业分割工具,将关键区域精准定位到医学图像中。

  3. 诊断
    结合经过校验的关键区域与医学证据,形成最终判断。

CARE 不再要求医生相信一个无法解释的结论,而是建立一条可以被检查、验证与审计的推理路径。

模型不只是说:

“这张影像可能提示某种疾病。”

它还可以进一步呈现:

  • 检查了哪个解剖结构;

  • 疑似异常位于什么位置;

  • 哪个区域影响了最终判断;

  • 诊断依据来自哪些证据。

这对医疗行业至关重要。

临床 AI 不能只依赖准确率数字,还必须让医生、医院和监管方理解建议是如何产生的。

CARE 正在尝试把医学影像推理,从黑箱预测转化为可追溯的循证流程。


Dragon Copilot 正升级为统一临床助理

在 HIMSS 2026 大会上,微软宣布 Dragon Copilot 已经不再只是一个独立的语音与文书工具。

它正在成为与 Microsoft 365 Copilot 连接的统一 AI 临床助理。

系统可以覆盖更广泛的工作,包括:

  • 临床文书生成;

  • 医患对话记录;

  • 医疗信息检索;

  • 工作流辅助;

  • 后续任务处理;

  • 跨 Microsoft 应用协作。

微软还在 3 月推出 Copilot Health,允许用户导入来自美国超过 5 万家医院和医疗机构的电子病历。

方向越来越清晰。

微软并不打算把医疗 AI 限制在某一款孤立的诊断模型中。

它正在构建连接临床数据、医疗文书、循证信息与办公应用的智能体层。


行业焦点正在从“大模型”转向“智能体”

医疗 AI 的核心问题正在发生变化。

过去几年,行业关注的是基础模型:

  • 哪个模型更懂医学?

  • 哪个模型医学考试分数更高?

  • 哪个模型能读取更长的病历?

  • 哪个模型参数规模更大?

现在,讨论逐渐转向智能体:

  • 哪个智能体负责听诊?

  • 哪个智能体负责写病历?

  • 哪个智能体检查质量?

  • 哪个智能体检索临床证据?

  • 哪个智能体跟进患者?

  • 哪个智能体连接医院信息系统?

IDC 在 CHIMA 2026 上的一项观察直指问题核心:

AI 技术供给正在开始超过医院的承载能力。

最大的缺口已经不再只是模型能力。

而是医院现有系统能否连接、管理和治理智能体。

许多医院仍然运行着相互割裂的信息系统:

  • 电子病历系统;

  • 检验系统;

  • 医学影像系统;

  • 药房平台;

  • 医保系统;

  • 患者管理平台;

  • 行政数据库。

即使一款智能体单独表现很好,如果它无法获取正确数据、无法把结果送回正确流程,或者缺乏明确权限控制,也无法创造真正价值。

新的挑战不只是训练更聪明的模型。

而是建设一套能够承载完整智能体生态的基础设施。


被忽视的问题:算力与 Token 由谁来提供?

当医疗智能体开始规模化运行,另一个问题随之出现:

支撑这些智能体的算力与 Token,究竟由谁来提供?

看看医疗工作流正在产生的调用规模。

讯飞门诊系统已生成超过 10 万份病历。

腾讯的 14 项医疗智能体能力可能同时运行在不同流程中。

微软 CARE 在观察、定位和诊断的每一个阶段,都需要调用专业模型和工具。

一次普通诊疗可能触发:

  • 语音识别;

  • 病历生成;

  • 质量审查;

  • 风险分析;

  • 指南检索;

  • 证据验证;

  • 随访规划。

每一个环节都需要消耗模型调用、上下文、工具操作与算力资源。

随着智能体数量增加,Token 消耗并不只是简单线性增长。

多个智能体可能重复读取同一份病历、交换中间结果、调用不同专家模型、验证输出,并对失败步骤进行重试。

总成本可能快速增长。

大型医院或许可以签署专属云合同,并建设独立算力基础设施。

但中小医院、基层医疗机构、独立医疗开发者和区域医疗服务商,面对的是完全不同的经济账。

它们需要先进模型,却未必有能力分别维护多家供应商的接口、账户和计费体系。

这时,Token 供给与结算层便成为医疗 AI 基础设施的一部分。


Unexhub:重构全球 Token 分发逻辑

Unexhub 正在尝试解决这一问题,重新组织 Token 与 AI 算力的聚合、分发、定价和结算方式。

Unexhub 的核心定位,是建设面向 Token 聚合与分发的算力交易中心。

平台将分散的模型和算力供给,与企业、开发者和 AI 应用产生的需求进行匹配。

对医疗智能体生态而言,这种能力越来越重要。

一家医院可能同时需要:

  • 病历生成模型;

  • 长上下文分析模型;

  • 医学影像多模态模型;

  • 低成本分类模型;

  • 高风险复核模型;

  • 主通道异常时的备用资源。

如果每个模型都分别签合同、开账户、做接口和处理账单,成本和运营复杂度都会迅速上升。

统一的分发平台可以简化这一过程。


即时结算,缓解中小供应商资金压力

对于中小 AI 供应商而言,最困难的问题之一是现金流。

它们立即提供算力或模型服务,但收入可能要在数周甚至数月之后才能结算。

在等待回款期间,供应商仍然要持续支付:

  • GPU 成本;

  • 电力费用;

  • 带宽费用;

  • 云端基础设施费用;

  • 运维成本;

  • 工程团队费用。

即使市场需求不断增长,过长的账期也可能让小型参与者承受巨大压力。

Unexhub 的结算清算系统,强调更快速的调用收益结算:

算力被调用后,相应收益可以更快回到供应端。

这种机制能够缩短供应商从提供资源到获得回款之间的时间。

对中小模型厂商和独立算力供应商来说,快速结算不仅是体验优化。

它可以降低周转资金压力,并帮助供应商继续投入更多算力。

Unexhub 还计划推出面向早期供应商的激励体系。

符合条件的早期参与者,在推广阶段可享受最低 0 佣金的平台费率,帮助其以更低成本进入分发网络,并尽可能保留初期调用收益。

这项权益的价值不只是“免费入驻”,还包括:

  • 降低渠道获取成本;

  • 更快连接企业需求;

  • 缩短收益结算周期;

  • 保留更大的定价测试空间;

  • 更快回收基础设施投入。


API 竞价,让 Token 价格回归市场

Unexhub 另一项值得关注的机制,是计划中的 API 竞价模式。

传统 API 价格通常由模型厂商固定。

无论算力是否空闲、资源是否充足,用户都只能按照公开价格调用。

Unexhub 尝试引入更加市场化的定价机制,让 API 价格根据供需变化动态调整。

在需求较低的时段,开发者可以用更低价格调用闲置资源。

当需求进入高峰,供应商则有机会通过提供稀缺资源获得更高收益。

这对供需双方都有价值。

对开发者与企业而言:

  • 可获得更低的闲时 API 价格;

  • 拥有更多模型路由选择;

  • 减少对单一供应商的依赖;

  • 更容易控制大规模 Token 开支;

  • 可根据预算灵活安排非紧急任务。

对供应商而言:

  • 提升闲置算力利用率;

  • 更容易触达全球需求;

  • 在高峰期获得更合理收益;

  • 更快结算调用分成;

  • 通过统一渠道降低分发成本。

这并不只是一次传统意义上的打折活动。

它尝试将部分 Token 定价过程,从供应商固定报价转化为市场供需驱动。


从固定 API 价格走向 Token 交易市场

医疗智能体的发展,让这种转变更加迫切。

医院并不是每一个任务都需要调用最昂贵的模型。

诊疗流程中的不同环节,对模型能力有不同要求:

医疗任务更合适的模型类型
语音转写快速低成本模型
病历生成均衡型语言模型
病历质量检查专业审核模型
医学影像分析医疗多模态模型
高风险诊断复核高端推理模型
诊疗指南检索搜索与循证智能体
患者随访低成本对话模型

如果所有环节都调用最昂贵的旗舰模型,医疗智能体的成本将很难持续。

更合理的系统,应根据以下因素动态选择资源:

  • 临床风险;

  • 任务复杂度;

  • 延迟要求;

  • 数据类型;

  • 使用预算;

  • 当前算力供给;

  • 服务可用性。

这正是 Token 交易市场的逻辑。

平台并不是只提供一个更便宜的 API。

而是让每一种任务都尽可能匹配更加合适的模型与算力资源。


Unexhub 的模型聚合与毫秒级路由

Unexhub 计划聚合不同规模的 AI 供应商。

平台可以接入海内外主流模型,包括:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 其他大型和专业 AI 模型。

个人开发者、初创团队和中小企业提供的优质模型与算力,也可以加入平台供应体系。

更加多元的资源供给,为企业与开发者提供了更广泛的选择。

通过毫秒级路由,请求可以根据不同因素被分配到合适通道:

  • 模型能力;

  • 价格;

  • 延迟;

  • 可用性;

  • 地区;

  • 任务类型;

  • 服务质量。

对医疗应用而言,路由能力尤其重要。

低风险的行政任务,不应该消耗与高风险诊断复核相同的昂贵资源。

统一路由层可以帮助医院和开发者,将高端算力留给真正需要它的场景。


按实际 Token 使用量结算

Unexhub 支持按照实际 Token 消耗进行计费。

开发者不必为了测试多个模型,分别预购大量固定套餐。

对独立开发者、初创公司和中小医疗机构而言,这种方式更加灵活:

  • 可以从较小预算开始;

  • 同时测试多款模型;

  • 根据真实需求逐步扩容;

  • 避免预付资源闲置;

  • 切换模型时不必重建整套应用。

在部分供应条件和推广资源方案下,实际 API 调用价格最低可达到市场标准价格的约 30%。

对高调用量的医疗智能体工作流来说,这会明显改变经济账。

真正节省的并不只是一次调用的几分钱。

同样的预算可能支持:

  • 更多诊疗咨询;

  • 更多病历生成;

  • 更多循证检索;

  • 更多独立复核;

  • 更长的智能体流程;

  • 向更多基层医疗机构部署。

当然,医疗机构不能只看价格。

任何服务于医疗场景的 Token 或 API 平台,都必须同时接受以下方面的评估:

  • 数据安全;

  • 隐私保护;

  • 模型可靠性;

  • 地区合规;

  • 审计能力;

  • 服务连续性;

  • 故障恢复;

  • 权限控制。

只有当低成本与安全、稳定的基础设施相结合时,价格优势才真正具有意义。


两场产业变革正在汇合

当医疗智能体与 Token 市场放在一起观察时,未来图景会变得更加清晰。

讯飞四智能体、微软 CARE、腾讯全栈方案与 Dragon Copilot,正在将诊疗流程改造成一条协同运行的 AI 工作流。

病历生成、质量审查、诊断支持、循证检索和患者随访,开始由不同专业智能体分别承担。

医生可以减少重复文书工作,将更多精力重新投入临床判断与患者沟通。

但这条自动化流水线要想持续运行,底层算力就不能中断。

它需要:

  • 稳定的模型访问;

  • 足够的 Token 容量;

  • 动态资源路由;

  • 透明的定价机制;

  • 更快的结算体系;

  • 多家备用供应商;

  • 可靠的基础设施。

Unexhub 正在建设这一底层能力。

医疗智能体企业在重新定义诊疗流程应该完成什么。

Unexhub 则试图重新定义支撑这条流程的算力应当如何供给和结算。

一边重写工作。

一边重写基础设施。

两者交汇的地方,才是真正的产业变革。


挑战依然存在

方向已经清晰,但转型不会轻松完成。

医院仍然需要回答一系列困难问题:

  • 现有系统能否承载多个智能体?

  • 数据能否安全地跨部门流动?

  • AI 诊断建议是否足够可解释?

  • 每一个临床结论能否追溯到证据?

  • AI 推荐错误时由谁承担责任?

  • 基层机构能否承担足够的 Token 成本?

  • 高峰期供应商能否保证稳定算力?

医疗 AI 不能像普通办公聊天机器人一样部署。

它运行在高风险环境中,任何错误都可能影响患者健康与安全。

因此,未来的医疗智能体生态必须同时具备多层治理机制:

  • 医生人工监督;

  • 证据可追溯;

  • 权限控制;

  • 独立验证;

  • 模型监控;

  • 安全数据处理;

  • 稳定的 Token 与算力供给。

技术正在从单点智能迈向协同工作流。

治理与基础设施也必须同步前进。


写在最后

诊室里的流水线正在被重新设计。

科大讯飞利用四类专业智能体,将病历生成、质量控制、诊断提示与循证检索连接在一起。

腾讯正在医院、医保与药企场景中建设模块化的医疗智能体能力。

微软 CARE 正把医学影像推理变成可以审计的证据链。

Dragon Copilot 则向统一临床助理演进,并融入更广泛的生产力生态。

这些系统解决的是一个问题:

医疗智能体应该做什么?

Unexhub 关注的是另一个问题:

支撑这些智能体的模型、算力、Token、价格与结算,应该如何被高效组织?

它通过 Token 聚合与分发,将全球 AI 供给与企业和开发者需求进行连接。

更快速的结算机制,可以帮助供应商更快获得调用收益。

计划中的竞价体系,可以让开发者在闲时获得更低价格,也让供应商在高峰需求中获得更合理回报。

统一模型接入与毫秒级路由,则可以让不同任务匹配不同资源。

医疗智能体正在重写诊疗流程。

Unexhub 正在重写其底层的 Token 分发逻辑。

一个改变医疗工作如何完成。

一个改变支撑这些工作的算力如何抵达。

医疗 AI 的未来,不会只由模型决定。

它还将由协调诊疗流程的智能体,以及让这些智能体持续运行的基础设施共同决定。