
A 2.8-trillion-parameter model now has a public repository. This piece weighs what self-hosting buys against what it costs, covers the middle options, and lists licence checks and vendor questions.
Suppose your legal team asks that internal documents not be sent to an outside provider, and someone on your team forwards the repository link for a 2.8-trillion-parameter model with the note "we can just download it and run it ourselves." It sounds like an answer. In practice it opens several new questions: who buys the hardware, who applies the security patches, and who picks up the phone at 3am when the service stops.
This article uses one model that recently gained a public repository as the example for a decision. We have not assessed its quality, compared prices, or recommended using or avoiding it. What we want to help you produce is a one-page rationale you can defend to an executive or a procurement committee: why the organisation runs a model itself, or why it does not.
What the repository says, and what it does not prove
On 12 November 2026 we opened the moonshotai organisation page on Hugging Face. It lists Kimi-K3.1-Pro at 2.8T parameters, shown as updated 7 days earlier, with 33.6k downloads. Kimi-K3, also 2.8T, was last updated on 2 September.
The Kimi-K3.1-Pro model page describes an open-weight model that accepts text and images, with 2.8T total parameters of which 104B are active per request, in a Mixture-of-Experts design. The licence it names is the Kimi K3.1 Pro License, a custom licence from the developer. The file list shows 96 safetensors files and a repository of about 1.56 TB. The card recommends vLLM and SGLang as inference engines but states no GPU or memory requirement.
The outlet letsdatascience.com reported that Moonshot launched an API for this version on 23 October 2026 and planned to release the weights on 4 November. That is a single source, so we use it only as a rough timeline.
None of this shows that the model works well on your organisation's Thai-language documents, because the benchmark results on the card are the developer's own and we do not use them. None of it shows that the licence covers your use. A download count does not mean any organisation runs the model in production. Some reports also describe the licence name and the model's status differently; we could not resolve that, so the file inside the repository is the one you must read yourself.
Open-weights versus open source, and the licence file
Open-weights means the model's numerical files can be downloaded. Open source, as the term is used in software, means a licence that grants the right to use, modify and redistribute under published conditions. These answer different questions. Being able to download a model does not tell you whether you may use it commercially.
The LICENSE file for Kimi-K3.1-Pro is headed Kimi K3.1-Pro License. We read it only to see its shape, not as lawyers. As we read it, the file contains conditions tied to revenue and user numbers, exemptions for internal use and for use through channels the developer designates, and a disclaimer of liability. We do not reproduce the thresholds or definitions here because interpreting them is a job for your counsel, and you should read the full text before commercial use.
The checks we would put in any licence due diligence for a model:
The licence name and version match the release you will actually run, since earlier releases from the same developer may use different text.
Any conditions tied to revenue, user counts or the form of service, and whether your organisation or its affiliates fall within them.
Whether internal use and giving outside parties access to outputs are treated alike.
Any requirement to display the model's or developer's name in your product.
The liability disclaimer, and who bears the loss when an output is wrong.
The licences of components used alongside it, such as the serving software.
A dated copy of the licence file as downloaded, because a repository file can change later.
One more check is often forgotten: who the licensee is. If a contractor runs the model for you, the licence attaches either to the contractor or to your organisation, and that decides whose revenue and users are counted.
What self-hosting actually buys
First, control over where data travels. If the model runs on a network you define and you control outbound traffic, prompts and answers do not pass through another provider's systems. That matters for documents you do not want anyone outside the network to see. Application logs, search indexes and backups remain yours to manage.
That control has to be demonstrable, not assumed because the server sits in your own machine room. Expect to show an outbound firewall rule for the model hosts, a list of who holds administrator rights, and logs that can be audited later. Those three are what an auditor will ask for before accepting that data does not leave.
Second, no per-token fee. Hardware and operators are a fixed cost you pay whether or not anyone calls the model. The break-even point depends on your real request volume. We have no one's prices to offer, and nobody should hand you a figure without seeing your workload.
Third, no vendor deprecating the version you rely on. You can pin the files you have for as long as you like. The price is that you decide when to move to a newer release and live with the consequences of not moving.
Costs that are not on the hardware quote
Start with memory. The arithmetic here is ours, not the developer's. We take the parameter count of 2.8 trillion from the listing and choose 8-bit precision for illustration. At one byte per parameter the weights alone are about 2.8 TB; at 4-bit, about 1.4 TB. That is before the memory the serving software needs for context. On accelerators with 80 GB each, 8-bit weights alone need at least 35 devices (2,800 GB divided by 80 GB).
Real deployments differ. The repository size of about 1.56 TB sits closer to the 4-bit figure, because the card describes MXFP4 weights. Ask the hardware seller what the true footprint is once you include the context length you need and the number of concurrent users. As a general property of Mixture-of-Experts models, all experts must be held in memory, so the 104B active figure says more about speed than about how much you must provision.
Hardware also has a contract term. A machine sized for one model may not fit the next one. Ask before buying when the repair and lease contracts end and what you will run on the machine then.
People are the cost the quote leaves out.
Serving software such as vLLM or SGLang is a server that accepts requests, so there is an operating system, accelerator drivers, container images and many libraries to patch. The party that publishes weights does not normally take on the job of maintaining your servers. A named person must track vulnerability notices and decide when to update. Files in the terabyte range pulled into an internal network are also a supply-chain path: confirm the source repository, record a hash for each file, and note who downloaded them and when.
Then there is re-evaluation. The organisation page we opened shows Kimi-K3 updated on 2 September and Kimi-K3.1-Pro updated around early November, roughly two months apart. If you run the model yourself, each new release raises the question of whether to move, and the answer should come from a set of your own real tasks that you re-run every time, not from the developer's results table.
The last cost is the 3am incident. When the server stops, memory fills, or responses slow abnormally, no provider takes the ticket. Whoever wakes up must understand the hardware, the network, the serving software and the calling application, plus a repair contract with a stated response time.
The middle options
A regional provider's API with data location fixed in the contract. You operate no servers, but you must read the contract terms on storage location, log retention, subprocessors, and the provider's rights to use your data.
Dedicated tenancy from a cloud provider. The model runs in an account or environment set aside for your organisation, on hardware the provider operates. Ask which layers the provider patches and which remain yours, because services divide that duty differently.
A smaller open-weights model. The same organisation's listing runs from 16B to 2.8T, so sizes exist. Hardware needs fall sharply, but the patching, evaluation and licence work remain, and you must show with your own test set that the smaller model is enough for the task.
If lock-in to one provider is the worry, our article on multi-cloud versus vendor lock-in covers how to weigh it, and the same reasoning applies to choosing a model provider.
What to ask a vendor who proposes to self-host for you
Who patches the operating system, drivers and inference server, within how many days of a vulnerability notice, and under which contract clause?
Who evaluates the model after each new release, with whose test set, and who decides whether to move?
Who holds the hardware and repair contracts: the vendor, or your organisation?
Who is the licensee of the model, and whose revenue and users does the licence count?
Where are prompts, logs and uploaded files stored, for how long, and who can read them?
Who answers out of hours, and what response time is agreed?
When the contract ends, what does the organisation receive: weights, configuration, runbooks, accounts and keys?
The last question is the one people ask least and regret most at handover. We covered what should change hands when a custodian changes in our article on vendor continuity due diligence.
Limits we should state plainly
If your real requirement is only that data stay in a region you choose, running a model of this size yourself is far more than the problem needs. A contract with a provider located in that region is the simpler path, it can be checked on paper, and it needs no hardware team. Where data may legally sit is a question for your organisation's counsel. We wrote about region choice in our article on region and data residency; this article does not say what any law requires.
We have also not run this model, and we did not read the licence as lawyers. Everything above comes from reading public pages on 12 November 2026, and those pages can change.
An objection we often hear
The objection is that per-token API fees add up to a large sum and running it yourself is cheaper. That may be true, especially with a high and steady volume of requests through the day. We have no one's prices to argue against, and we do not think a reader should trust a figure that has not been run against their own workload. What we ask is that the self-hosted side be counted in full: hardware idle overnight, out-of-hours cover, spare machines, and the time the team spends evaluating each release. Compare that with the monthly request volume you have actually measured. If the two sides still land close together, the remaining difference is who carries the risk when the system stops.
Where to start
Write the real requirement in one sentence, for example "data must stay in one region" or "nothing leaves the corporate network." That sentence decides whether a middle option is enough.
Collect a set of real tasks that can be used to measure any model, without waiting to pick one.
Pilot on an option that has a contract behind it, and record actual request volume for a few months.
Send the licence file, with its download date, to counsel for review.
Set the condition that will bring you back to the self-hosting question, such as request volume reaching an agreed level or a change in the provider's terms.
Summary
Before agreeing to run a model yourself, write down the name of the person who answers the call at 3am when the model server stops, and the name of the person who evaluates after each new release. If you cannot write both names, a middle option whose contract already assigns those duties is usually the safer place to begin, and counsel still has to read the model's licence in every case.
If you would like to talk through the details of a project, you can reach us at 088-983-9386 or [email protected]. Our office is in Bang Kapi, Bangkok. For organizations still drafting a scope of work, we are glad to arrange a conversation to understand the problem first, without needing complete documents. And if we find that the problem is not yet at the point of needing a new system, we will say so plainly, because starting a project at the wrong time costs more than waiting another quarter.
FAQ: Frequently Asked Questions about This Article
A collection of questions and answers to help you better understand the content of this article.
Tags:

