Qingdao Port Launches 35bn-Parameter Xinggang Model for 80 Scenarios

The Analysis

Qingdao Port released Xinggang, a port and shipping large model with 35 billion parameters, covering more than 80 business scenarios through three agents for berth, equipment and yard plus 12 digital planning modules. Berth planning time collapsed from three to four hours down to one to three minutes, with accuracy above 90 percent. Throughput capacity rose 10 percent, rehandling fell 5 percent and energy cost dropped 8 percent. In the first eight months the port moved 513 million tonnes of cargo, up 3.1 percent, and 23.57 million TEU, up 6.9 percent.

The headline everyone will repeat is 35 billion parameters. That number is the least useful thing on the page, and it is the one a competitor will quote to pretend it is behind. I am a systems person, and systems people do not price models in parameters. We price them in the hours they remove and the percentage points they lift, because a model that does not move a number on an operations dashboard is a very expensive way to consume electricity. So let us convert the headline into the units that actually decide whether this is worth copying.

Berth planning is a scheduling problem, and scheduling is where a large port loses or saves the most invisible money. Before the model, a planner spent three to four hours assembling a berth plan for the coming shift or the coming day. That plan decides which ship goes to which berth, in what order, with which cranes, and it feeds everything downstream: equipment allocation, yard stacking, truck and rail pickups. Three to four hours of planning time is not just labour. It is latency. A plan produced three hours before it is needed cannot react to a ship that arrives early, a crane that goes down, or a batch of containers that is not where the yard map says it is. By the time you can re-plan, the window has moved.

Cut that to one to three minutes and the whole character of the operation changes, because now the plan can be rebuilt during the shift, not just before it. A berth plan that can be regenerated in minutes turns a port from a schedule-follower into a schedule-setter. When a ship arrives early, the planner can re-optimise berths in real time and squeeze an extra call into the day. When a crane fails, the re-plan can shift the crane assignment before the delay compounds. The value here is not the three hours of planner time saved. Planners are not that expensive. The value is the ability to re-plan at all, because re-planning is what absorbs variance, and variance is what wrecks throughput. Take the three hours saved at, say, the productivity of a busy berth, and the removed latency is worth a multiple of the labour saved. That multiple is why the model exists, and it has nothing to do with 35 billion parameters.

Now look at the accuracy figure, above 90 percent, because accuracy is where a scheduling model earns or loses its keep. A berth plan is only as good as its inputs, and the inputs are estimates: how long the ship will take to discharge, how fast the cranes will actually run, whether the yard stacks are as the system believes. A model that plans to above 90 percent accuracy is not claiming to be perfect. It is claiming to be reliably better than a human planner working under time pressure, and that is a much more demanding and much more useful claim.

The reason is that planning errors compound. A berth plan that is wrong by an hour causes a cascade of wrong equipment, wrong truck appointments and wrong yard moves, and each of those wrong decisions costs money and time. A plan accurate to 90 percent cuts the cascade roughly in proportion. So the 90 percent is not a quality badge. It is a variance-reduction statement, and variance reduction is the whole game in port operations.

Translate the throughput and rehandling numbers into a common currency, because they are being stated separately and they belong together. Throughput capacity up 10 percent means the port can move a tenth more boxes through the same physical footprint. Rehandling down 5 percent means fewer wasted container moves, where a box is picked up and put back because it was stacked in the wrong order. Every rehandling move is pure cost, no revenue, plus extra equipment wear and extra energy. So the 5 percent rehandling reduction is a direct cost cut, and it compounds with the 10 percent throughput gain, because freeing up equipment and yard space from rehandling is part of how you get the throughput. The energy cost down 8 percent rides on the same logic: less wasted movement means less fuel and less electricity per box. Read three percentages as one mechanism, and the mechanism is better planning reducing wasted motion, which shows up as both more capacity and lower cost.

Scale it, because the port's own numbers give you the base. Eight months, 23.57 million TEU, up 6.9 percent. A 10 percent throughput capacity gain on a base that large is not a rounding error. Take the full-year figure implied by eight months, roughly 35 million TEU annualised, and 10 percent of that is 3.5 million TEU of additional capacity from a model, without a new berth, a new crane or a new hectare of yard. Compare the cost of a new berth, which in a large Chinese port runs into the hundreds of millions of dollars and takes years to permit and build, with the cost of training a model on data the port already owns. The model is orders of magnitude cheaper per unit of capacity, which is why every major port will eventually have one. The interesting question is not whether, but who gets the compounding benefit first.

Here is the overlooked point, and it is the one the sales pitch never mentions. A port model eats historical operating data, and its ceiling is set by how many years of that data the port has. A port with a decade of digitised berth, equipment and yard records can train a model that captures the port's real patterns, including its weird corners. A port with three years of clean data, or with data living in three systems that do not talk to each other, will train a model that is confident about the easy cases and useless on the hard ones, which are exactly the cases where a planner needs help.

So the differentiator between a big port and a smaller one will not be whether it has a model. Both will. It will be the accuracy of the plan on the hard days, and that accuracy rides on data depth and data cleanliness. A small port that buys the same software on thin, fragmented data will get a plan that looks smart on the demo and disappoints on a stormy Tuesday. The model is only as good as the pile of records behind it, and the pile takes years to build and years to clean.

That is also why the 12 digital planning modules and the three agents are the substance behind the parameter count. The agents for berth, equipment and yard are not one general brain. They are three specialised planners, each solving a distinct sub-problem with its own data and its own objective. Berth planning optimises ship-to-berth assignment against arrival and departure windows. Equipment planning optimises crane and vehicle assignment against the berth plan. Yard planning optimises stacking against the throughput the other two expect. The reason this is the right architecture is that a single general model trying to plan all three at once would be solving a problem with too many degrees of freedom, and the solution would be unstable, changing its mind when any input twitched. Three models in a defined order inherit stability from their sequence, because each one solves a bounded problem and hands a clean output to the next. That structural choice, not the parameter count, is what makes the plan usable in a real control room.

Now the counter-case, because a port model has a specific failure mode and it is worth watching for. A model trained on historical data learns the port as it was. If the port's traffic mix changes, or a new berth opens, or a carrier switches its rotation, the historical pattern no longer describes the present, and a model that has learned the old pattern can be confidently wrong about the new one. The accuracy above 90 percent was measured against the conditions it learned. Under new conditions, it will drift down until it is retrained.

So the real operational discipline is not buying the model. It is maintaining the feedback loop that retrains it as conditions change, and staffing the control room with people who can tell when the model's plan looks wrong and override it. A port that installs the model and trusts it blindly will find it degrades quietly. A port that treats it as a strong default with a human override will keep it sharp. That maintenance discipline is invisible in the launch announcement and is where the long-run value sits.

For a shipper, none of this is abstract, because it lands on your berth window and your cut-off. A port with a fast, accurate planning system gives you two practical gifts. It can hold a tighter berth window, which means less waiting at anchorage and a more reliable departure, and it can absorb a late container that would otherwise miss the ship, because re-planning is cheap enough to do during the shift. Both improve your effective transit time without any change to the advertised sailing. So when you choose a port or negotiate a service through one, the port's planning maturity is a real variable in your transit variance, even though it never appears in a rate quote. A port with Qingdao's class of system can promise a tighter effective window than one still planning by hand, and that promise is worth money to a shipper whose delivery dates matter.

There is a queue-hiding point inside the berth plan that nobody puts on a slide, and it is where the real saving lives. A port is a set of queues stacked on top of each other, and the berth is the first queue, the one that sets the clock for every queue behind it. Berth time drives equipment demand, equipment demand drives yard demand, yard demand drives the truck and rail gates. When the berth plan is built three hours before it is used, every downstream queue inherits that plan's blind spots at the same moment, because the plan was frozen when the information was thin.

A plan built in three minutes runs on fresher information for every one of those queues, so the model is not just fixing the berth. It is feeding cleaner inputs to equipment, yard and gate at the same time. That is the multiplier a parameter count can never describe, because the parameter count is a property of the model while the queue effect is a property of the schedule the model writes.

Notice what the accuracy number is actually measuring, because it is easy to read it as a quality score when it is really a variance measure. In any queueing system, throughput collapses not when the average service time rises a little, but when the variance rises a lot. A port that plans to the average, with high variance around it, will clog even at moderate load, because a plan that is wrong by an hour in either direction chokes the berth, then the cranes, then the yard. A plan generated in minutes and accurate to above 90 percent is a variance-reduction machine, and in a queue, cutting variance buys more throughput than cutting the mean. That is why the same physical port can absorb more boxes from a planning change alone, and it is the mechanical reason the 10 percent throughput number is achievable without new steel in the water.

Consider the cost side the same way, because rehandling is just variance you paid for in container moves. A rehandle is a container picked up and set back down because it was stacked in the wrong sequence for its next move. Every rehandle consumes a crane hour, a yard slot, a fuel burn and a slice of equipment life, and returns nothing. Five percent off rehandling is five percent off a pure-loss activity, and because rehandles also occupy the yard and the cranes, cutting them frees the same resources that then serve the 10 percent throughput gain. Energy cost down 8 percent is the fuel and electricity bill of those removed moves, so the three percentages are not three separate wins. They are one mechanism, better planning producing less wasted motion, showing up once as capacity and twice as cost.

Put an order of magnitude on the maintenance question, because this is where a model quietly becomes an expense instead of an asset. A model trained on a decade of clean records will keep its accuracy as long as the port's traffic mix stays roughly stable, and a port's traffic mix does change, sometimes fast, when a carrier reroutes, a neighbouring terminal opens or a new berth shifts the pattern. Each of those events makes the historical pattern stale, and a stale pattern produces confident errors on exactly the cases the planner cannot easily catch. So the true operating cadence of a port model is a retraining schedule plus a human override layer, and the model's accuracy claim is only valid inside the corridor those two maintain. A port that skips the retraining gets a model that decays on a curve nobody watches, and decay in a planning model does not announce itself. It just shows up as slightly worse berth windows that everyone blames on the weather.

Roughly speaking, the value of a port model shows up in three places, and a shipper can watch all three. Berth planning time, hours to minutes. Planning accuracy, the plan versus the actual outcome. And the throughput and rehandling numbers, which are the financial summary of both. If a port publishes these, you can compare ports on their planning maturity rather than on their crane count. If a port publishes only the parameter count, it is telling you it would rather talk about its model than its results, and that is itself a signal about how the results look. The ports that can prove the hours and the percentage points will lead. The ones reciting parameters will follow and pretend they are not.

There is one more arithmetic layer worth putting on, because it decides where the model pays off first inside a port. Take a berth plan with four vessels slated for the day and a weather window closing in the afternoon. A hand plan made at dawn is fixed through the morning and, if the weather comes early, it costs a whole vessel's delay. A model that can re-plan in two minutes lets the planner pull that last vessel forward before the window shuts, recovering the call. One recovered call a day, at the call's margin, over a year is a number that dwarfs the model's cost several times over, and it happens in the exact conditions where a port's reputation is made or broken. The best use of a port model is not the easy day. It is the bad day, and the bad day is the only day a shipper remembers.

So the reading I would take away is this. The 35 billion parameters is a marketing unit. The operating units are minutes off the planning cycle, percentage points of accuracy, ten percent of throughput, five percent of rehandling and eight percent of energy, and those five numbers are what a port should be judged on. The next competitive gap between big and small ports will not be the presence of a model. It will be the depth of the data underneath it and the discipline of the humans over it, because those are what determine whether the plan holds on the days the port is under stress. And stress, not the calm Tuesday, is when you find out whether the model was worth building.

↑ Back to the key points

— By Ravi Chandran