“Only by anchoring to its own DNA and focusing on real pain points can technology truly leave the lab and become productive force that solves problems.”
In the wave of smart-transport transformation, travel large models often fall into the paradox of "advanced technology yet hard deployment": algorithm parameters keep breaking records, yet they still struggle to solve real problems like "rural buses running empty, idle charging piles, and lagging highway-accident response." China Mobile's self-developed "Jiutian Chuanliu" travel large model, built on its telecom DNA, precisely cracks this paradox. Its core logic is not chasing algorithmic extremes, but returning to the first principle of mobility governance and using gene-fitted solutions to break the wall between technology and scenario.
01 The Essential Contradictions and Causes of Travel Large-Model Deployment
The deployment deadlock of travel large models essentially departs from the first principle of "dynamic matching efficiency of people - vehicles - roads - time," and three core contradictions cause a total loss of supply-demand matching ability.
(1) The Industry's Essential Contradiction:
The triple fracture of supply-demand matching: traditional travel models have fatal shortboards across the whole "perception - prediction - decision" chain:
Data blind spots cause "incomplete supply-demand perception": traditional models rely on navigation-APP user-reported data, with rural and remote-area coverage only 60%, unable to capture whole-domain travel demand. Taking rural Miyi, Sichuan as an example, in 2024 lack of basic data caused bus schedules to mismatch farming-season travel, with an empty-load rate as high as 40%.
Lagging prediction causes "insufficient matching timeliness": relying on historical-data modeling, response latency to sudden conditions like accidents or temporary controls reaches 1-3 seconds, unable to support dynamic matching. Guangdong's Xinbo Expressway 2024 data shows that due to prediction lag, average road-recovery time after accidents exceeded 90 minutes.
Disconnected decisions cause "failure of supply-demand linkage": data for driving, cycling, and metro are fragmented, and output advice is isolated and unlinked. Hangzhou's Binjiang District 2024 charging-pile layout, not linked to metro-passenger data, placed 30% of devices in low-demand areas with under-35% utilization.
(2) Root Causes of the Contradictions:
The genetic defects of traditional models: the core of these contradictions is not insufficient algorithm precision, but a gene mismatch between the model base and the travel scenario:
Innately insufficient data sources: over-reliance on "user-initiated authorized data" naturally misses key samples like non-APP users and rural groups, causing structural blind spots in supply-demand perception.
Statically ossified technical architecture: using traditional machine-learning models, unable to fuse dynamic road-condition data in real time, prediction ability decays over time, hard to fit the dynamicity of travel scenarios.
Isolated and closed application design: functions split by single travel mode, with no "demand-side - supply-side" linked closed loop formed, decision advice cannot land as actual optimization actions.
02 AI Application Deployment: A Penetration Path From "Single-Point Optimization" to "Systemic Governance"
Jiutian Chuanliu's deployment did not take the "big and all" sprawl route, but focused on the core pain points of transport governance, forming a "layered penetration, on-demand fit" pattern — most evident in government-enterprise cooperation.
(1) Urban Traffic: From "Passive Congestion Cure" to "Active Dispatch"
In core urban areas like Chengdu High-Tech Zone and Hangzhou's Binjiang District, the model's core value is cracking the hard problems of "rigid signal timing and blind facility layout." By analyzing multi-modal travel data of cycling, metro, and driving, the model outputs refined decision advice: Chengdu optimized signal timing at 200+ intersections based on the model, cutting morning-peak congestion duration by 20%; Hangzhou adjusted charging-pile layout based on crowd-flow patterns, raising utilization from 45% to 65%. More notable is its "natural-language interaction" capability — grassroots traffic managers need not master complex algorithms; by directly asking the "Chuanliu Zhixing" app "what is the evening-peak passenger peak at a certain business district," they get data support and decision advice, greatly lowering the threshold of technology deployment.
(2) Highway Emergency: From "After-the-Fact Handling" to "Before-the-Fact Warning"
On transport trunk lines like Guangdong's Xinbo Expressway, the model focuses on the pain point of "lagging accident response," building a "prediction - dispatch - handling" closed loop. Relying on 5G dedicated-network low-latency traits, the model can predict traffic-flow peaks and high-risk accident sections 2 hours ahead, guiding clearance vehicles and rescue equipment to deploy early, raising emergency-response speed by 40% and cutting dispatch cost by 25%. This capability shows even more value in emergency scenarios: in Sichuan's Jiuzhaigou flood disaster of 2025, drone base stations carrying the model quickly restored communication, while the model real-time-simulated optimal routes for rescue vehicle fleets, safeguarding 1,200 person-times of rescue.
(3) Cross-Domain Extension: From "Transport Governance" to "Industry Empowerment"
The model's value is breaking beyond transport, penetrating linked industries like cultural tourism and commerce. In scenic-area management, analyzing tourists' spatiotemporal movement trails can early-warn overcrowding risk; in business-district planning, based on dwell time and consumption-correlation data, it provides a basis for shop layout and promotion activities. This extension is not a tech crossover, but captures the essence that "travel data is consumption-demand data" — what communication base stations capture is not only "where people move," but "what latent demand people have," providing a brand-new data entry for industry digitalization.
Jiutian Chuanliu's core competitiveness is converting China Mobile's "whole-domain network coverage" DNA into a travel-matching solution, achieving precise fit across the data, technology, and application layers.
(1) Data Layer: Base-Station Whole-Domain Perception Completes the Matching Foundation
Relying on China Mobile's infrastructure advantage of 2.8 million+ 5G base stations, the model achieves 95% geographic coverage of "city - rural - remote areas," processing 1-billion-level location data daily, and needs no user-initiated authorization — communication base stations naturally capture people's spatiotemporal movement trails, solving the sample-bias problem at the source. This "whole-domain penetrability" data capability is the core barrier traditional models cannot replicate.
(2) Technology Layer: Dynamic Algorithms Lift Matching Timeliness
Using an ultra-large-scale dynamic graph neural network, fusing "2018-2024 historical data + real-time base-station signals" with tens-of-billions of parameters, it achieves sudden-road-condition response latency < 500 ms and total traffic-flow prediction error < 3.7%. In Guangdong's Xinbo Expressway application, the model can predict high-risk accident sections 2 hours ahead and guide early deployment of rescue equipment; the synergy of dynamic-graph algorithms and communication data lets "before-the-fact prediction" replace "after-the-fact response."
(3) Application Layer: Multi-Modal Fusion Achieves the Matching Closed Loop
Breaking data-fragmentation barriers, it integrates multi-dimensional data of driving (traffic flow), cycling (shared-bike dispatch), and metro (passenger peaks) to output "full-chain matching advice." For example, Hangzhou's Binjiang District in 2025, based on model advice, linked charging-pile layout with metro-passenger data and optimized bus stops with cycling-connection points, raising bus on-time rate. This "demand - supply" linked application design turns decision advice into real governance results.
The sustainable deployment of AI large models hinges on building an ecosystem of "technology supply - demand fit - value sharing." Jiutian Chuanliu, relying on China Mobile's capability mid-platform AaaS + ecosystem, has walked a "low-cost, replicable" inclusive path.
On the technology-supply side, the model is integrated into the "Juzhi Agent Development Platform," forming synergy with 5 major industry large models including finance and marketing, supporting "zero-code" rapid application building. This means SMEs need no AI team; by simply calling model capabilities through the platform they can solve their own travel-related needs, confirming the scale effect of technology supply.
On the government-enterprise collaboration side, the model adopts a "data desensitization + capability output" cooperation mode, both safeguarding user privacy and meeting government governance needs. In cooperation with Chengdu's Traffic Management Bureau and Hangzhou's Urban Management Bureau, the model does not deliver raw data directly, but outputs decision products like "signal-optimization plans" and "charging-pile layout advice" — this "result-oriented" cooperation lowers the government's data-security concerns and accelerates deployment. By 2024, related projects had driven over 4.2 million yuan in revenue and won the "Outstanding Case Award" at the Global AI for Good Summit, achieving a win-win of economic benefit and social recognition.
03 Lessons from the Case: Let Large Models Become the Productivity That Solves Real Business Pain Points
Jiutian Chuanliu's success is essentially China Mobile converting its core DNA of "communication-network coverage" into the AI competitiveness of "travel-data perception." This transformation path offers three insights for traditional enterprises:
First, data assets must be "used deeply" rather than "piled massively."
The model did not blindly chase data scale, but focused on the uniqueness of communication data being "whole-domain, continuous, objective," building a differentiated advantage — more deployment value than simply integrating multi-source data.
Second, scenario fit must "stick to pain points" rather than "chase hot spots."
From urban congestion cure to highway emergency, every scenario aimed at the real hard problems of transport governance, replacing concept hype with quantifiable results like "20% drop in congestion duration" and "38% lower accident rate" — this is the key to the model's government-enterprise recognition.
Third, ecosystem building must "lower thresholds" rather than "build barriers."
Through carriers like the "zero-code" platform and capability mid-platform, SMEs and grassroots governments can all use AI capabilities at low cost; such an inclusive ecosystem has far more vitality than a closed technical system.
Of course, model deployment still faces challenges: how to balance data use and privacy protection, how to cope with base-station data bias under extreme weather, and how to achieve cross-operator data collaboration — all are topics requiring sustained cracking. Jiutian Chuanliu's practice offers an important lesson for traditional enterprises' AI transformation: only by anchoring to their own DNA and focusing on real pain points can technology truly leave the lab and become the productive force that solves problems.
