skip to content

Model Sourcing

Where the model you ship comes from: a provider's API or open weights, the licence those weights carry, and where you run them. Interviewers ask which you would pick and whether you may ship it.

on this pageshow

explore

questions

17

What separates a closed hosted model from open weights for a team building an LLM feature?

level: juniorimportance: must knowfreq 55%

answer

  1. who holds the weights
  2. an API versus a checkpoint file
  3. the provider decides when it changes
  4. one checkpoint, many possible hosts

basics

~20 s

A closed hosted model is reached only through an API its provider or a licensed partner runs, and the provider can change it; open weights are a checkpoint the team holds, to tune, inspect and serve where it chooses.

solid answer

~40 s

With a **closed hosted model** you send requests to an API run by the provider or a partner it licenses: you never see the weights, you get only the outputs and controls the API exposes, and the provider decides when the model is updated or retired. With **open weights** you hold the checkpoint itself, so you can tune every parameter, read full log-probabilities and internal activations, and have any compatible host, or your own hardware, serve it. The usual trade is capability and convenience against control: closed frontier models often lead on the hardest tasks and need no infrastructure from you, while open weights give up some of that lead for inspection and portability. What the weights' licence allows and what running them costs are separate questions.

go deeper

for a junior

Recall the one-line split: closed means an API onto weights you never see, open means you hold the checkpoint. Then name one thing each side gives you.

for a middle

Explain what the API withholds, such as activations, full log-probabilities and the tuned weights from hosted fine-tuning, and what holding a checkpoint makes possible in return.

for a senior

Tie the choice to a concrete requirement of the feature, like calibrated scores, a model that must not change, or a tiny team with weeks to ship, rather than to ideology.

for a principal

Frame it as control against convenience and frontier capability, and keep licence, hosting and cost as separate decisions the team makes after this one.

## Two ways to get a model Every LLM feature rests on a sourcing call: where does the model come from? There are two broad answers. - A **closed hosted model** is a model whose trained parameters (its **weights**) are never released. You reach it through an API that the provider runs, or that a partner such as a cloud platform runs under the provider's licence. You send a request, you get a response, and everything between is the provider's. - **Open weights** means the trained parameters are published as a downloadable **checkpoint**: the files that hold every weight of the network, plus its tokenizer and configuration. Whoever holds those files can load them into a compatible serving stack and run the model. The distinction is about **who holds the weights**, not about quality, price or licence. An open checkpoint can carry a restrictive licence, and a closed model can be cheap. ## What a closed hosted model gives and withholds A closed model gives a team a working endpoint on day one: - **No infrastructure**: no accelerators to rent, no serving engine to tune, no on-call rota for the model itself. - **Often the strongest capability**: the best closed models usually lead the best open ones on the hardest reasoning, coding and long multi-step tasks, although the size of that lead varies by task and over time. - **Managed extras**: many providers add hosted fine-tuning, tool calling and structured output to the API. It also withholds things, because you only ever see the API: - You cannot read the model's **internal activations**, and many APIs return **log-probabilities** only for a few top tokens per position, or none at all. - **Hosted fine-tuning**, where offered, runs the provider's methods on the provider's servers; the tuned weights never leave. - The provider decides **when the model changes**: an update, a new default version or a retirement happens on its schedule, not yours. ## What open weights give and cost Holding the checkpoint changes what is possible: - **Full tuning**: you can update every parameter, or train adapters, with whatever method you choose, and keep the result. - **Full inspection**: you can read the probability of every token in the vocabulary and the hidden states at every layer. - **Portability**: the same checkpoint can be served by several unrelated hosts or on your own hardware, so the model does not disappear if one supplier does. - **A frozen reference**: the files you downloaded do not change unless you change them. The cost is that somebody has to **run it**. Even when a managed host serves it, the team now chooses the host, checks how the host serves the checkpoint, and owns more of the evaluation work. And the best open checkpoints have usually trailed the closed frontier on the hardest tasks. ## Side by side | Question | Closed hosted model | Open weights | |---|---|---| | Who holds the weights? | The provider | The team (and anyone else who downloads them) | | Who can serve it? | The provider and its licensed partners | Any compatible host, or your own hardware | | Tuning | The provider's hosted options, if any | Any method, every parameter, result kept | | Log-probabilities and activations | Limited or none | Full | | When does it change? | When the provider decides | When the team decides | | Capability at the frontier | Usually ahead | Usually somewhat behind | | Infrastructure work | None for the model itself | Some, depending on where it runs | ## What this choice does not settle Picking closed or open is the first call, not the only one. Each of these is its own question: 1. **The licence on open weights.** Downloadable does not mean free for any use; whether a team may ship what it plans is a matter of reading the licence. 2. **Where open weights run.** A shared managed endpoint, a dedicated one, rented accelerators and your own racks each hand back different responsibilities. 3. **What running them costs** against paying per token, which is a separate break-even calculation. 4. **Where data may go**, including what a provider retains, which belongs with data-handling policy. ## How interviewers probe it A screening question usually wants the core distinction stated cleanly: closed means you rent access to a model you cannot see or keep; open means you hold the model and take on running it. Strong answers add the consequence that matters for the feature at hand, such as "we need calibrated scores, so we need log-probabilities the API may not give us", or "we are three people shipping in a month, so we start with an API". Weak answers treat open weights as automatically free to use, assume a closed model never changes, or assume open weights require buying hardware.

  • Can a team fine-tune a closed hosted model?
    Often, through the provider's hosted fine-tuning, but on the provider's terms: the team picks from the methods and settings the provider offers, never receives the tuned weights, and can serve the tuned model only where that provider serves it. With open weights the team chooses the method, tunes any parameter it likes and keeps the result.
  • Why might a team pick a closed model even when an open checkpoint could do the task?
    Because there is nothing to run. A small team, a prototype with unproven demand or a task near the capability frontier can gain more from a working endpoint on day one, with no accelerators or serving work, than from control it does not yet need, provided it accepts that the provider sets the pace of change.

  • Open weights mean the model is free to use for anything.
  • A closed model behind an API never changes once you start using it.
  • Open weights are far behind closed models on every task.
  • Choosing open weights means you must buy and run your own GPUs.
  • A closed model can never be fine-tuned.
open as a page

Why is an open-weight model you can download not necessarily open source, and what would its licence have to drop to qualify?

level: juniorimportance: must knowfreq 40%

basics

~20 s

Downloadable weights tell you about access, not rights. Open source means anyone may use, modify and redistribute for any purpose; many weight licences add use policies, user-count caps or non-commercial terms that fail that test.

open as a page

What can a team do with open weights that a closed API offering hosted fine-tuning still does not allow?

level: middleimportance: must knowfreq 50%

basics

~20 s

Open weights let a team tune every parameter with any method and keep the result, read log-probabilities over the whole vocabulary for any text, and read or change internal activations; hosted fine-tuning offers only the provider's methods, on its servers.

open as a page

For open-weight models, when would you use a shared serverless endpoint rather than a dedicated managed endpoint, and what do you give up?

level: middleimportance: must knowfreq 42%

basics

~20 s

A shared serverless endpoint suits an unmodified catalog checkpoint with modest or bursty traffic and no replicas to run. You give up custom weights, a pinned serving build, reserved capacity and replica-level isolation, which a dedicated managed endpoint gives back.

open as a page

How do permissive, custom, research-only and non-production weight licences differ in what a team may put into production?

level: middleimportance: must knowfreq 34%

basics

~20 s

Permissive licences allow production with notice duties. Custom licences allow it under conditions such as use policies, user thresholds and attribution. Research-only and non-commercial licences bar commercial use. Non-production licences permit evaluation but need a separate grant to deploy.

open as a page

How big is the capability gap between frontier closed models and the best open weights, and how has it moved?

level: middleimportance: should knowfreq 45%

basics

~20 s

The best closed models usually lead on the hardest reasoning, coding and long multi-step tasks, with open weights trailing by roughly months to a year; the gap has narrowed since early chat models, widens at frontier releases and varies by task.

open as a page

After fine-tuning an open-weight model, which hosting rungs can still serve your checkpoint, and what does each rung need from it?

level: middleimportance: should knowfreq 36%

basics

~20 s

Rented accelerators and own racks serve whatever your serving engine can run. A dedicated managed endpoint serves uploaded full weights of an architecture the host supports. A shared endpoint usually serves only its catalog, sometimes with adapters on a catalog base.

open as a page

An open-weight model's licence requires a separate grant above a monthly-active-user count: what do you check about how and when that count is measured?

level: middleimportance: should knowfreq 24%

basics

~20 s

Check whose users count (one product, the whole company, affiliates), what counts as an active user, whether it is measured at a fixed date or continuously, and whether the grant above it is guaranteed or at the publisher's discretion.

open as a page

A contract-clause classifier built on a closed frontier model now needs calibrated confidence scores; how does that requirement push the sourcing choice?

level: seniorimportance: should knowfreq 30%

basics

~20 s

Calibration needs a per-label probability that can be checked and rescaled on held-out data. If the closed API returns too few log-probabilities or none, the team is left with costly sampling or unreliable stated confidence; open weights give the full label distribution.

open as a page

Why does building on one closed model lock a team in even when every provider accepts the same request shape, and how does an open checkpoint change that?

level: seniorimportance: should knowfreq 45%

basics

~20 s

Lock-in lives in prompts, examples, output parsing and eval thresholds tuned to one model's behaviour, not in the request format, so leaving means retuning and re-validating them; an open checkpoint served by several hosts lets a team change supplier without changing the model.

open as a page

A regulated customer needs a fine-tuned open-weight model with no shared tenancy: which hosting rung would you choose, and what isolation does each rung give?

level: seniorimportance: should knowfreq 30%

basics

~20 s

First pin down which layer "no shared tenancy" covers. Shared serverless fails at the batch. A dedicated managed endpoint isolates replicas and usually accelerators; single-tenant rented hardware adds the physical server; only hardware you or the customer run excludes the host's operators.

open as a page

Moving open-weight serving from a managed endpoint to rented GPUs and then your own racks, which layers of the stack become yours to patch?

level: seniorimportance: should knowfreq 34%

basics

~20 s

On rented GPUs you take the guest operating system, the accelerator driver, the compute runtime and its libraries, the serving engine and the container images. Your own racks add firmware, management controllers, the network fabric and failed-hardware replacement.

open as a page

When a team fine-tunes, quantises or distils from an open-weight checkpoint, which result inherits the base licence, and what can an output clause add?

level: seniorimportance: should knowfreq 28%

basics

~20 s

A fine-tune and a quantised build modify the weights, so they carry the base licence's terms. A student distilled only from outputs is bound as far as the licence's output and derivative clauses reach, and some licences reach it explicitly.

open as a page

Why can serving an open-weight model behind your own API and shipping its weights inside a customer's on-premises appliance carry different licence obligations?

level: seniorimportance: should knowfreq 26%

basics

~20 s

Many weight licences attach their heaviest duties to distribution: licence copies, notices, attribution and passing the use policy on. Serving keeps the weights with you; shipping them in an appliance makes the customer a recipient those terms must reach.

open as a page

When is starting on a closed model with a planned exit to open weights a sound plan, and what must a team keep portable?

level: principalimportance: should knowfreq 35%

basics

~20 s

It is sound when shipping soon matters more than control, open weights can plausibly meet the task within the planning horizon, and the team keeps its own task evals, prompts, output contracts and data in forms no single provider owns.

open as a page

Leading a platform team that serves open weights for several product teams and one air-gapped customer, how far down the hosting ladder would you go, and why?

level: principalimportance: should knowfreq 24%

basics

~20 s

Decide per workload: keep teams on managed endpoints unless a custom checkpoint or tenancy rule forces self-run accelerators. The air-gapped customer alone forces hardware with no outbound path, served from one standard self-run stack updated through a reviewed import path.

open as a page

What licence review would you require before a product commits to an open-weight checkpoint, and what should trigger that review to run again?

level: principalimportance: nice to knowfreq 20%

basics

~20 s

Review each checkpoint, not each family: record the licence and referenced policy as accepted, its category, and each use, threshold, derivative and output clause against your plans and deployment shape. Legal signs off; re-review on any change.

open as a page