苹果 PrismML 让 iPhone 跑 27B,腾讯 Hy3 单卡扛 295B——AI 算力交易中心的跨域调度逻辑要重写

On July 14, Tencent Hunyuan announced low-bit quantized versions of its 295-billion-parameter flagship model, Hy3. The model was compressed into 1-bit and 4-bit formats, packaged in GGUF and connected to the llama.cpp ecosystem.

Apple Runs 27B on an iPhone, Tencent Fits 295B on a Single GPU: Why Compute Trading Centers Must Rewrite Their Cross-Domain Scheduling Logic

English Version

On July 14 and 15, 2026, the AI industry witnessed two events that appeared unrelated but, when viewed together, pointed to the same structural shift.

On July 14, Tencent Hunyuan announced low-bit quantized versions of its 295-billion-parameter flagship model, Hy3. The model was compressed into 1-bit and 4-bit formats, packaged in GGUF and connected to the llama.cpp ecosystem.

A model that once depended on a multi-GPU cluster could now be deployed with significantly fewer hardware resources—and, under suitable configurations, potentially on a single inference-class accelerator.

The original BF16 weights occupied nearly 600GB. The 1-bit version was reduced to approximately 85.5GiB, representing a compression ratio of about 6.7 times.

More importantly, the compressed model reportedly preserved much of its long-context understanding capability. Performance declined only moderately in more demanding areas such as agent workflows and code generation.

One day later, on July 15, PrismML released Bonsai 27B.

Fine-tuned from Qwen 3.6 27B, Bonsai was designed to run natively on an iPhone equipped with 12GB of memory while preserving most of the original model’s capabilities.

At 16-bit precision, the original model required roughly 54GB of storage. Even a conventional 4-bit version occupied about 18GB. Bonsai’s 1-bit edition reduced that requirement to only 3.9GB.

On an iPhone 17 Pro, the model reportedly achieved an inference speed of approximately 11 tokens per second.

American software engineer Tim Carambat described the significance in unusually direct terms: the impact of this kind of on-device model compression may ultimately prove more important than another round of frontier-model benchmark competition.

One company compressed a 295B model toward single-accelerator deployment.

Another placed a 27B model inside a smartphone.

Within 48 hours, the practical ceiling of edge-side reasoning was raised twice.

This was not merely a model-compression story.

It was a warning that the traditional distribution logic of computing power was about to be rewritten.


AI Inference Is Moving From Cloud-Centric to Cloud-Distributed

For much of the past five years, the dominant narrative in the AI industry was simple:

Models would continue getting larger, and computing power would continue moving into the cloud.

Parameter counts increased. Context windows expanded. Inference workloads became more complex. Multi-GPU servers and centralized data-center clusters appeared to be the inevitable foundation of advanced AI.

Hy3 illustrates that old logic clearly.

Its original BF16 weights were close to 600GB, making multi-card deployment almost unavoidable. A 27B model requiring 54GB at 16-bit precision was equally unrealistic for a consumer smartphone.

Low-bit quantization is now overturning those assumptions.

This is no longer a gradual improvement measured in a few percentage points. It is an order-of-magnitude compression in model size and memory requirements.

Tencent reduced nearly 600GB to about 85GB.

PrismML reduced approximately 54GB to 3.9GB.

As compression ratios move beyond tenfold in selected scenarios, the idea of running capable models on edge devices is shifting from research demonstration to deployable reality.

Industry forecasts have suggested that a growing share of AI inference workloads will migrate toward edge devices by 2026. Releases such as Bonsai 27B and quantized Hy3 could accelerate that trend.

When a smartphone can run a 27B model locally, and a single accelerator can support a model with hundreds of billions of parameters, many tasks that once had to be sent to a centralized cloud environment may be handled closer to the user.

These tasks may include:

  • Multi-step reasoning

  • Structured tool invocation

  • Private document analysis

  • Long-context workflows

  • Local coding assistance

  • Persistent agent loops

  • Real-time multimodal interaction

The more capable the device becomes, the less frequently every request needs to return to a distant cloud cluster.

Cloud computing will not disappear.

But the cloud will no longer be the only place where meaningful AI inference takes place.


The Distribution Model Is Changing, Not Simply Shrinking

It would be misleading to conclude that edge-side inference will cause cloud demand to collapse completely.

The real change is more structural.

AI computing power is moving from a cloud-centralized model toward a distributed architecture spanning three layers:

device, edge and cloud.

Routine, privacy-sensitive and latency-sensitive tasks can increasingly remain on the device.

Tasks requiring regional coordination, caching or moderate-scale acceleration can move to nearby edge nodes.

Large-scale reasoning, frontier-model inference, model training and highly complex agent orchestration will continue to rely on centralized cloud infrastructure.

The question is therefore no longer:

Should this task run in the cloud?

The more important question is:

At this exact moment, should this part of the task run on the device, at the edge or in the cloud?

That decision changes dynamically according to:

  • Model size

  • Available memory

  • Device temperature

  • Battery level

  • Network latency

  • Privacy requirements

  • Token price

  • Task complexity

  • Service availability

  • Required output quality

This is the new economic and technical foundation of AI inference.

And it creates a major challenge for computing power trading centers.


The “Computing Power Bank” Must Move Beyond Cross-Regional Scheduling

On July 8, the China Academy of Information and Communications Technology launched an ecosystem initiative centered on two concepts:

  • The computing power supermarket

  • The computing power bank

The computing power bank introduces an operational framework built around the deposit, metering, withdrawal and credit-based use of computing resources.

Its goal is to enable flexible access to computing power through cross-regional and cross-period scheduling.

The original design logic is easy to understand.

If computing capacity is insufficient in eastern China, idle resources in western regions can be scheduled.

If daytime demand is too high, workloads can be shifted to lower-demand periods at night.

This is a natural extension of the “East Data, West Computing” era, in which computing power is primarily understood as large-scale infrastructure concentrated inside data centers.

However, Bonsai 27B and quantized Hy3 introduce a new variable:

edge-side computing power.

Once smartphones, personal computers, workstations, automobiles, smart devices and local enterprise servers gain meaningful inference capabilities, computing power can no longer be defined only as GPUs installed in data centers.

It also includes an enormous number of distributed inference nodes located in the hands of users and inside local business environments.

These nodes are fragmented.

They are intermittently available.

Their performance varies.

They are widely distributed across geographic locations.

Yet their combined computing capacity could eventually exceed that of any single centralized cluster.

The scheduling logic of the computing power bank must therefore evolve from simple cross-regional scheduling to genuine cross-domain scheduling.

It must coordinate not only among data centers, but across the device, edge and cloud layers.


From Data-Center Capacity to Billions of Distributed Inference Nodes

The current computing power bank concept encourages small and medium-sized enterprises to deposit idle computing resources.

That is an important first step.

However, the idle capacity of a small enterprise server is fundamentally different from the fragmented inference capacity of hundreds of millions—or potentially billions—of consumer devices.

A business server may remain online for long periods and offer relatively predictable performance.

A smartphone may become unavailable at any time because:

  • The user starts another application

  • Battery power falls below a threshold

  • The device enters power-saving mode

  • Network conditions change

  • The device is moving

  • Temperature limits reduce performance

  • The user withdraws permission

This means edge-side computing power cannot simply be added to a traditional resource pool.

It requires a new operating system for compute circulation.

The system must understand not only how much computing power exists, but also:

  • Where it is located

  • When it is available

  • How stable it is

  • What model it can support

  • What privacy boundary applies

  • How much energy it will consume

  • What service quality it can guarantee

  • How much the resource should cost

The rise of on-device reasoning is therefore creating a structural change on the supply side.

In the past, hundreds of data centers supplied most commercial AI computing power.

In the emerging model, millions of enterprises and vast numbers of devices may contribute smaller units of inference capacity simultaneously.


Unexhub Is Positioned at This Turning Point

Unexhub is building a computing power trading center focused on the aggregation and distribution of APIs and tokens.

Its goal is to reshape the way global model capacity is connected, priced, routed and settled.

The platform brings together AI resources across different capability and supply levels, including model ecosystems associated with Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as quality services offered by independent developers and small and medium-sized providers.

This structure becomes increasingly important as computing power moves from centralized clusters toward fragmented, multi-layer supply.

The computing power bank can establish the institutional framework.

Unexhub can provide the market circulation layer.

If the computing power bank is similar to the financial infrastructure of the computing economy, Unexhub functions more like the trading venue and clearing network.

One manages standards, deposits, credit and resource scheduling.

The other manages aggregation, transactions, price discovery, routing and settlement.

Together, a “bank plus market” architecture could absorb and redistribute fragmented computing resources more efficiently.


Real-Time Settlement Makes Small Suppliers More Competitive

The first requirement of a distributed computing market is settlement efficiency.

Traditional model and cloud distribution channels often operate with long billing cycles. Small suppliers may provide large volumes of service but still wait weeks—or longer—to receive payment.

This creates substantial working-capital pressure.

Unexhub’s real-time settlement mechanism links revenue more closely to actual token consumption. As model calls are generated, suppliers can receive their corresponding share according to platform settlement rules.

This can significantly shorten the revenue recovery cycle and reduce the financial burden of participating in the market.

Eligible early-stage suppliers may also receive 0% platform commission benefits, allowing them to retain more of their initial revenue while testing demand and building a customer base.

The value of this policy is not limited to saving a service fee.

For smaller suppliers, lower commission and faster settlement can mean:

  • Less cash locked inside long billing cycles

  • Lower customer acquisition pressure

  • Faster reinvestment in infrastructure

  • More room to offer competitive pricing

  • Easier entry into global distribution channels

This makes fragmented suppliers more capable of competing with larger platforms.


API Bidding Turns Idle Capacity Into Market Value

The second requirement is dynamic pricing.

Traditional cloud pricing assumes that computing resources have relatively stable value throughout the day.

Distributed inference does not follow that logic.

The value of a model endpoint changes with:

  • Real-time demand

  • Available capacity

  • Network latency

  • Model quality

  • Geographic location

  • Energy cost

  • Reliability

  • Time of day

Unexhub’s API bidding mechanism allows participating resources to compete according to current supply and demand.

During periods of lower utilization, developers may access suitable model capacity at significantly lower prices. For selected routes and usage periods, the effective API cost may fall to approximately 30% of comparable market pricing.

During high-demand periods, suppliers with reliable capacity and better service quality may earn higher returns.

This creates a more balanced market:

Developers benefit from lower-cost access.

Suppliers gain the opportunity to monetize idle capacity and receive a premium when demand rises.

Rather than treating lower prices as a temporary discount, API bidding turns price flexibility into a permanent market mechanism.


Millisecond-Level Routing Connects Device, Edge and Cloud Resources

Dynamic pricing alone is not enough.

Users should not need to manually compare dozens of models, providers, regions and endpoints before sending every request.

Unexhub uses intelligent routing to match workloads with suitable model channels according to price, latency, availability, quality and user-defined requirements.

As device-side and edge-side computing resources enter the market, this routing layer becomes even more valuable.

A future request might be split across several environments:

  • Private preprocessing runs locally on the device

  • Retrieval runs on a nearby edge node

  • Complex reasoning is sent to a cloud model

  • Verification is completed by another provider

  • The final answer is assembled locally

The user experiences one request.

Behind the scenes, the workload may cross several computing domains.

Millisecond-level routing allows the platform to continuously choose the most appropriate path without requiring developers to maintain separate integrations for each provider.

Developers can pay according to actual token usage.

Suppliers can connect through a unified channel and receive revenue based on completed calls.

This is how fragmented computing power becomes a usable market rather than a collection of disconnected resources.


Three Ways the Scheduling Logic Must Be Rewritten

The rise of capable on-device models requires computing power trading centers to rebuild their scheduling logic across three dimensions.

1. Scheduling Granularity Must Move From Clusters to Devices

Traditional scheduling focuses on clusters.

A scheduler allocates a workload to a GPU server, an availability zone or a data center.

In a cross-domain architecture, the scheduling target may be:

  • A smartphone

  • A personal computer

  • An enterprise workstation

  • A local edge server

  • A vehicle computing platform

  • A regional GPU cluster

  • A hyperscale cloud environment

The device becomes the smallest potential supply unit.

This does not mean every phone should automatically be turned into a public computing node. User consent, security, battery consumption and data privacy must remain fundamental constraints.

But technically and economically, the scheduling system must be capable of recognizing device-level resources.

The computing power bank’s deposit-and-withdrawal model currently focuses mainly on enterprise resources.

The next stage will be extending that logic toward trusted and permission-based terminal capacity.


2. Billing Must Move From Card-Hours to Tokens and Outcomes

Card-hour and core-hour billing remain useful for infrastructure procurement.

But they are too coarse for a distributed inference market.

Two devices may consume the same amount of time while delivering completely different volumes and qualities of output.

Tokens provide a more portable unit of measurement.

They allow computing resources from different chips, regions and providers to be compared through a common consumption framework.

As token-based computing products, token stores, token factories and computing power supermarkets emerge, tokens are increasingly being treated as resources that can be measured, priced, distributed and traded.

In this environment, card-hours describe the cost of owning infrastructure.

Tokens describe the value of consuming intelligence.

Future settlement systems may go even further by combining token volume with:

  • Output quality

  • Response latency

  • Task success rate

  • Energy consumption

  • Model precision

  • Service reliability

Token billing will not eliminate every infrastructure metric, but it is likely to become the common language connecting suppliers and users.


3. Scheduling Must Move From Static Planning to Real-Time Decision-Making

Traditional cross-regional scheduling often relies on relatively stable capacity plans.

End-side computing power is far less predictable.

A phone may be available now and unavailable thirty seconds later.

An edge node may have spare capacity in the morning and become overloaded during a local event.

A cloud endpoint may suddenly reduce its price or experience a latency spike.

The scheduling system must continuously sense changing conditions and decide:

  • Which tasks should remain on the device

  • Which tasks should move to the edge

  • Which tasks require cloud inference

  • Which model should handle each stage

  • Which provider offers the best current price-performance ratio

  • Whether the task should be delayed, split or rerouted

This is no longer simple cross-regional scheduling.

It is cross-domain, real-time orchestration.


A New Infrastructure Is Beginning to Take Shape

Apple running a 27B model on an iPhone and Tencent compressing a 295B model toward single-accelerator deployment are not merely technical achievements.

They represent a shift in computing economics.

When inference moves from cloud centralization toward device-edge-cloud distribution, the logic of computing power trading must evolve with it.

Scheduling granularity must be redesigned.

Billing units must become more universal.

Pricing must become dynamic.

Settlement must become faster.

Routing decisions must happen in real time.

Tokens are emerging as the shared language of this transformation.

The computing power bank manages the deposit, withdrawal, credit and scheduling of resources.

Token factories standardize and process computing output.

Unexhub is building the market layer where APIs and tokens can be aggregated, routed, priced and settled.

These components are gradually forming a new infrastructure for device-edge-cloud collaboration.

The platform’s value is reflected on both sides of the market.

For developers and enterprise users:

  • Multiple model resources can be accessed through a unified interface

  • Billing is based on actual token consumption

  • Millisecond-level routing reduces manual switching costs

  • Selected routes may be available at prices as low as 30% of comparable market rates

  • Workloads can be matched dynamically according to price and performance

For suppliers:

  • APIs can be connected through a unified distribution channel

  • Usage revenue can be settled in real time or at high frequency

  • Eligible early participants may receive 0% commission benefits

  • Idle capacity can be converted into measurable revenue

  • Dynamic bidding creates additional earning opportunities during demand peaks

The significance of the two releases in July 2026 is therefore larger than the compression ratios themselves.

One model was compressed into a single-accelerator deployment range.

Another was compressed into a smartphone.

Together, they revealed the next question the industry must answer:

When intelligence can run everywhere, who will coordinate, price and settle the computing power behind it?


苹果在手机上运行27B,腾讯让295B迈向单卡部署:算力交易中心的跨域调度逻辑需要重写

中文版

2026年7月14日至15日,AI行业接连发生了两件看似独立、实则互为映照的事件。

7月14日,腾讯混元发布旗舰模型Hy3的低比特量化版本。

这款拥有2950亿参数的模型被压缩为1-bit和4-bit版本,以GGUF格式发布,并接入llama.cpp生态。原本高度依赖多卡集群的模型,在适当配置下,已经能够以更少的硬件资源部署,甚至逐渐接近单张推理级显卡可承载的范围。

Hy3原始BF16权重接近600GB,而1-bit版本被压缩至约85.5GiB,整体体积缩小约6.7倍。

更值得关注的是,在长文本理解等任务中,1-bit版本仍然保留了较多原始模型能力,仅在Agent任务、代码生成等高复杂度方向出现一定程度的性能下降。

一天之后,7月15日,PrismML正式发布Bonsai 27B。

这款模型基于Qwen 3.6 27B微调而来,目标是在保留大部分原始能力的同时,原生运行于拥有12GB内存的iPhone之上。

原始模型在16-bit精度下需要约54GB空间,传统4-bit版本仍然需要约18GB,而Bonsai的1-bit版本仅为3.9GB。

据公布的信息显示,它在iPhone 17 Pro上的推理速度可以达到约每秒11个Token。

美国软件工程师Tim Carambat对其意义给出了十分直接的评价:相比新一轮旗舰模型榜单竞争,这类端侧模型压缩技术可能对行业产生更加深远的影响。

一个把295B模型压缩至接近单卡部署。

另一个把27B模型塞进手机。

短短48小时内,端侧推理能力的上限被连续抬高了两次。

这不仅是一场模型压缩技术的突破。

它也意味着,过去围绕云端集群建立起来的算力分配逻辑,正在被重新改写。


AI推理正从“云端集中”走向“端云分布”

过去五年,AI行业一直遵循一条看似明确的发展路线:

模型越来越大,算力越来越集中于云端。

参数规模持续增长,上下文窗口不断扩展,推理流程越来越复杂。多卡服务器与大型数据中心集群,似乎成为运行先进AI模型不可替代的基础设施。

Hy3原始版本就是这一逻辑的典型代表。

其BF16权重接近600GB,多卡服务器几乎是标准配置。一个在16-bit精度下需要54GB空间的27B模型,同样不可能直接运行在普通手机上。

低比特量化正在推翻这些既有判断。

这已经不是几个百分点的渐进式优化,而是模型体积与内存需求的数量级压缩。

腾讯将接近600GB的权重压缩至约85GB。

PrismML则将约54GB的模型压缩至3.9GB。

当部分场景中的压缩比例超过十倍,“旗舰能力模型运行在端侧”正在从实验室演示变成可以实际部署的产品形态。

此前已有行业报告预测,到2026年,将有越来越多的AI推理任务迁移至边缘和终端设备。Bonsai 27B与Hy3量化版本的出现,很可能进一步加快这一趋势。

当手机可以本地运行27B模型,单张推理卡能够承载数千亿参数级模型时,许多过去必须上传至云端的任务,都有机会在更靠近用户的位置完成。

这些任务可能包括:

  • 多步骤推理

  • 结构化工具调用

  • 私密文件分析

  • 长上下文工作流

  • 本地代码辅助

  • 持续运行的Agent循环

  • 实时多模态交互

端侧能够完成的任务越多,每一次请求都必须返回远程云端的必要性就越低。

云计算不会消失。

但云端将不再是高质量AI推理唯一能够发生的地方。


算力需求不是简单减少,而是改变了分布方式

如果据此判断云算力需求会彻底崩塌,同样并不准确。

真正发生的变化,是算力需求结构的重组。

AI推理正在从单一的云端集中模式,转向覆盖三个层级的分布式体系:

端、边、云。

常规任务、隐私敏感任务和低延迟任务,可以更多地留在本地设备。

需要区域协同、缓存加速和中等规模计算的任务,可以交由附近的边缘节点。

大规模复杂推理、前沿模型调用、模型训练以及高强度Agent编排,仍然需要云端数据中心完成。

未来需要回答的问题,不再只是:

这个任务是否应该放到云端运行?

而是:

在当前这一时刻,这个任务的哪一部分应该运行在端侧、边缘节点或云端?

答案会随着多种因素实时变化:

  • 模型规模

  • 可用内存

  • 设备温度

  • 电池电量

  • 网络延迟

  • 隐私要求

  • Token价格

  • 任务复杂度

  • 接口可用性

  • 输出质量要求

这将成为AI推理新的技术基础与经济基础。

同时,也对算力交易中心提出了全新的挑战。


“算力银行”需要从跨区域调度升级为跨域调度

7月8日,中国信息通信研究院启动了围绕两类创新服务形态展开的生态共建计划:

  • 算力超市

  • 算力银行

其中,算力银行探索建立包括算力资源存入、计量、取用和授信在内的运行机制,并通过跨区域、跨周期调度,实现算力资源的灵活配置。

这一设计逻辑并不难理解。

东部地区算力不足,可以调用西部地区的闲置资源。

白天需求过高,可以将部分任务调度至夜间低峰期。

这是“东数西算”时代的典型思路。在这一体系中,算力主要被理解为集中在数据中心内部的大型基础设施。

但Bonsai 27B和Hy3量化版的出现,引入了一个新的变量:

端侧算力。

当手机、个人电脑、企业工作站、智能汽车、本地服务器和各类智能终端都具备一定推理能力时,算力就不能再被简单定义为数据中心里的GPU。

它还包括大量分布在用户手中和企业本地环境中的推理节点。

这些节点高度碎片化。

可用时间并不连续。

性能规格各不相同。

地理位置也极为分散。

但当它们被连接起来时,其总规模可能超过任何单一数据中心。

因此,算力银行的调度逻辑必须从传统的跨区域调度,升级为真正的跨域调度。

它所调度的不应只是不同地区的数据中心,还应覆盖端、边、云三个层级。


从几百座数据中心,走向海量分布式推理节点

当前的算力银行机制,鼓励中小企业将闲置算力资源接入共享体系。

这是非常重要的一步。

但企业服务器的闲置算力,与数以亿计、甚至未来可能达到数十亿台终端设备的碎片化推理能力,存在本质区别。

企业服务器通常可以保持较长时间在线,性能也相对稳定。

手机则可能因为多种原因随时退出资源池:

  • 用户打开了其他应用

  • 电池电量下降

  • 设备进入省电模式

  • 网络环境发生变化

  • 用户正在移动

  • 温度限制导致性能下降

  • 用户主动撤回授权

这意味着,端侧算力不能简单并入传统资源池。

它需要一套新的算力流通操作系统。

系统不仅要知道“有多少算力”,还要清楚:

  • 算力位于哪里

  • 何时可以使用

  • 稳定性如何

  • 能够运行什么模型

  • 适用何种隐私边界

  • 会消耗多少能源

  • 可以保证什么服务质量

  • 应该如何定价

端侧推理的爆发,正在推动算力供给侧发生结构性变化。

过去,商业AI算力主要由几百座大型数据中心提供。

未来,数以百万计的企业节点和海量终端设备,都可能同时贡献更小颗粒度的推理能力。


Unexhub恰好位于这一转折点

Unexhub正在打造一个面向API与Token聚合、分发和交易的算力交易中心。

其目标,是重塑全球模型资源的连接、定价、路由与结算方式。

平台汇聚不同规模、不同能力层级的AI资源,包括Gemini、ChatGPT、Claude、DeepSeek、GLM等主流模型生态,同时也接入个人开发者和中小型供应商提供的优质模型服务。

当算力从集中式集群走向碎片化、多层级供给时,这种平台结构的重要性会进一步提高。

算力银行负责建立制度框架。

Unexhub则负责搭建市场化流通层。

如果将算力银行理解为算力经济的金融基础设施,那么Unexhub更接近于算力资源的交易市场与清结算网络。

一个负责标准、存取、信用与资源调度。

另一个负责聚合、交易、价格发现、智能路由与收益结算。

“银行+市场”的双层架构,能够更有效地承接和重新分配碎片化算力资源。


实时结算,让中小供应商拥有更强竞争力

分布式算力市场首先需要解决的,是结算效率问题。

传统模型和云服务分发渠道通常存在较长账期。中小供应商即使已经完成大量调用,也可能需要等待数周甚至更长时间,才能收到对应收入。

这会带来明显的资金周转压力。

Unexhub通过实时或高频结算机制,让供应商收入与实际Token调用更加紧密地关联起来。模型每产生一笔有效调用,供应商便可以根据平台规则获得对应的调用分成。

这一机制能够显著缩短资金回笼周期,降低中小供应商参与市场时的现金流压力。

符合条件的早期入驻供应商,还可以享受阶段性的平台0抽佣权益,在测试市场需求、积累客户与建立服务口碑的过程中,保留更多实际收入。

这一权益的价值并不只是节省一笔平台服务费。

对于中小供应商而言,更低的抽佣和更快的结算意味着:

  • 减少资金被长期账期占用

  • 降低前期获客压力

  • 更快将收入重新投入基础设施

  • 拥有更灵活的价格调整空间

  • 以更低门槛进入全球分发渠道

这使碎片化供应商也有机会与大型平台展开竞争。


API竞价,让闲置算力拥有实时市场价值

分布式算力市场的第二个要求,是动态定价。

传统云计算定价往往默认算力在一天中的价值相对稳定。

但分布式推理并不符合这一逻辑。

一个模型接口的实时价值,会受到多种因素影响:

  • 当前市场需求

  • 剩余可用容量

  • 网络延迟

  • 模型输出质量

  • 节点所在地区

  • 能源成本

  • 服务稳定性

  • 使用时间段

Unexhub通过API竞价机制,让不同模型资源根据实时供需参与市场竞争。

在资源利用率较低的闲时,开发者可以自动匹配更具性价比的模型通道。对于部分线路和特定使用时段,实际API调用成本最低可达到同类市场价格的约3折。

在需求旺盛的高峰期,能够提供稳定容量和高质量服务的供应商,也可以获得更高收益。

这一机制为市场两端同时创造价值:

开发者能够以更低成本获得模型服务。

供应商则可以将闲置容量转化为收入,并在需求上升时获得合理溢价。

因此,低价不再只是临时优惠。

它成为由供需关系持续驱动的市场机制。


毫秒级路由,连接端、边、云资源

仅有动态定价还不够。

开发者不应该在每次发起请求前,都手动比较数十个模型、供应商、地区和接口。

Unexhub通过智能路由,根据价格、延迟、可用性、输出质量以及用户自定义要求,为任务匹配更合适的模型通道。

随着端侧和边缘算力进入交易体系,这一智能路由层会变得更加关键。

未来,一个完整任务可能被拆分到多个环境中执行:

  • 隐私数据预处理在本地设备完成

  • 信息检索交由附近边缘节点执行

  • 复杂推理调用云端大模型

  • 结果验证由另一家模型供应商完成

  • 最终内容回到本地组装

用户只发起了一次请求。

但在后台,任务可能已经跨越多个算力域。

毫秒级路由可以持续选择更加合适的执行路径,开发者无需为每一家供应商分别维护接口。

开发者按照实际Token消耗付费。

供应商通过统一渠道接入,并根据有效调用获得收益分成。

这正是把碎片化算力从“彼此孤立的资源”,转化为“可以使用的市场”的关键一步。


调度逻辑需要从三个方向重写

端侧大模型能力的提升,对算力交易中心的调度逻辑提出了三个层面的重构要求。

一、调度颗粒度需要从“集群级”下沉到“设备级”

传统调度面对的是集群。

系统会将任务分配给某台GPU服务器、某个可用区或某座数据中心。

在跨域算力体系中,可被调度的节点可能包括:

  • 智能手机

  • 个人电脑

  • 企业工作站

  • 本地边缘服务器

  • 智能汽车计算平台

  • 区域GPU集群

  • 超大规模云计算中心

终端设备可能成为算力供给的最小单位。

这并不意味着所有手机都应该自动变成公共算力节点。用户授权、数据安全、电池消耗与隐私保护,必须始终是基础约束。

但从技术与经济角度看,调度系统需要具备识别设备级资源的能力。

目前算力银行的“存入—取用”机制,主要面向企业级算力资源。

下一阶段,则可能逐步延伸至可信、可控且经过用户授权的终端算力。


二、计费单位需要从“卡时”进化到“Token与结果”

卡时和核时仍然适用于基础设施采购。

但对于分布式推理市场而言,这类计费方式过于粗放。

两台设备即使运行相同时间,能够输出的Token数量、响应质量和任务完成效果,也可能完全不同。

Token提供了一种更容易跨平台流通的计量单位。

它能够让不同芯片、不同地区和不同模型供应商的资源,在统一的消耗框架中进行比较。

随着词元集合店、词元工厂和算力超市等新形态出现,Token正在被赋予可计量、可定价、可分发和可交易的资源属性。

在这一体系中,卡时描述的是基础设施占用成本。

Token描述的则是智能服务消费价值。

未来的结算体系,还可能进一步结合:

  • 输出质量

  • 响应延迟

  • 任务成功率

  • 能源消耗

  • 模型精度

  • 服务稳定性

Token计费不会完全取代所有基础设施指标,但它有机会成为连接供应商与用户的通用语言。


三、调度决策需要从“静态规划”转向“实时动态”

传统跨区域调度,通常建立在相对稳定的算力规划之上。

端侧算力却具有高度不确定性。

一台手机此刻可能处于空闲状态,30秒后便可能无法继续参与推理。

一个边缘节点早上可能拥有大量闲置容量,到了本地活动高峰期便迅速过载。

一个云端接口可能突然降价,也可能出现延迟上升。

调度系统必须持续感知这些变化,并实时决定:

  • 哪些任务留在终端运行

  • 哪些任务转移至边缘节点

  • 哪些任务必须调用云端模型

  • 每个环节应该选择什么模型

  • 当前哪家供应商的性价比更高

  • 任务是否需要延迟、拆分或重新路由

这已经不再是简单的跨区域资源调配。

而是跨域实时编排。


一套新的算力基础设施正在成形

苹果在iPhone上运行27B模型,腾讯让295B模型迈向单卡部署,并不仅仅是模型压缩技术的突破。

它们代表的是算力经济运行方式的改变。

当AI推理由云端集中走向端、边、云协同,算力交易体系也必须随之升级。

调度颗粒度需要重新设计。

计量单位需要更加统一。

价格需要动态形成。

结算周期需要进一步缩短。

路由决策必须实时完成。

Token正在成为这场变化的共同语言。

算力银行负责资源的存入、取用、授信和调度。

词元工厂负责将算力产出标准化和产品化。

Unexhub则在建设一个能够聚合、分发、路由、定价与结算API及Token的市场层。

这些组件正在逐步拼合出一套服务于端、边、云协同的新型算力基础设施。

对于开发者和企业用户,Unexhub能够提供:

  • 通过统一接口接入多种模型资源

  • 根据实际Token使用量付费

  • 通过毫秒级路由减少手动切换成本

  • 部分线路最低可享受同类市场价格约3折

  • 根据价格与性能动态选择模型通道

对于模型与算力供应商,平台能够提供:

  • 通过统一渠道快速接入分发网络

  • 根据实际调用实时或高频结算收益

  • 符合条件的早期供应商享受0抽佣权益

  • 将闲置容量转化为可计量收入

  • 在需求高峰期通过竞价机制获得更多收益

因此,2026年7月这两次发布的意义,远远超过模型压缩比例本身。

一个模型被压缩至接近单卡部署。

另一个模型被压缩进手机。

它们共同提出了AI行业必须回答的下一个问题:

当智能可以运行在任何地方,谁来完成背后算力的调度、定价与清结算?