Hosted fine-tuning: which platforms, which models, which trade-offs

0:00 Hosted fine-tuning

In twenty twenty-six, a lot of teams woke up to an email saying the fine-tuning platform they built on was being wound down. Their fine-tuned models would keep running for a while, but new training would end. The teams that shrugged were the ones who had kept their data and evaluations portable. In this lesson you will map the hosted fine-tuning landscape, the formats and APIs, and the checklist that keeps you able to move.

0:33 Analogy: tailoring a suit

Here is an analogy. Hosted fine-tuning is like tailoring a suit at a department store. It is convenient, they have the machines, and the result fits well. But the suit stays in their store, and if the store closes the tailoring counter, you cannot take the pattern elsewhere. Self-managed fine-tuning on open weights is owning the pattern. Either way, keep your measurements, which are your data and your evaluation, in your own notebook.

1:05 Snapshot (verify)

Here is the snapshot, as of September twenty twenty-six, and please verify it. Google's Vertex AI offers supervised fine-tuning for Gemini two point five Pro, Flash and Flash-Lite, plus preference tuning, and tuning for selected open models. Amazon Bedrock customizes Amazon Nova models, including supervised fine-tuning, distillation and reinforcement fine-tuning starting with Nova two Lite. Microsoft Foundry fine-tunes the GPT four point one and GPT four o families, with other methods and models on varying access.

1:38 What changed

And the changes. OpenAI began winding down its self-serve fine-tuning platform this year. It closed to new organizations, is phasing out new jobs for existing customers, and keeps existing fine-tuned models available for inference until their base models are deprecated. Mistral's docs now place its self-serve fine-tuning API under legacy resources. Meanwhile, several inference platforms offer LoRA fine-tuning on open-weight models, often letting you download the adapter.

2:07 Portability

That makes portability the key strategic question. A fine-tune of a closed model lives only with that provider. You keep your data and evaluation, not the weights. A fine-tune of an open model, on your GPUs or a platform that exports adapters, gives you portable weights. Either way, keep your datasets, labeling guides, graders and evaluation harness provider-neutral, and moving becomes days of work, not months.

2:36 Formats

Formats differ. OpenAI-style chat format, also used by Azure, is a list of messages with roles system, user and assistant. Gemini on Vertex uses contents with parts, and the assistant role is called model, plus a separate system instruction. TRL uses prompt and completion. The lesson shows the same example in all three. Keep one canonical format and write small converters, so switching is a script, not a project.

3:06 Launching jobs

The lesson includes two illustrative job launches. On Vertex AI, the Google Gen AI SDK's tunings dot tune call takes a base model, a training dataset in Cloud Storage and a config with a display name and validation data, and you poll until it finishes. On Bedrock, boto3's create model customization job takes a job name, role, base model, customization type, S3 locations and hyperparameters, whose names differ per base model. Check each SDK's current docs before running.

3:40 Six-point checklist

Before choosing, run a six-point checklist. Which base models can you tune, and in which regions? Which methods? How is training data handled: retained, used for anything else, encrypted, region-bound? What does inference cost for custom models, and is dedicated capacity required? What happens when the base model is deprecated? And can you export weights or adapters?

4:05 Worked example: Riyadh e-commerce

A worked example. A Saudi e-commerce firm needs an Arabic intent classifier with data staying in the kingdom. They check which providers offer tuning and inference in an in-kingdom region, find their preferred closed model is not tunable there, and choose LoRA on an open-weight model using in-kingdom cloud GPUs. Their adapters are portable, and their canonical data and harness mean they can compare a hosted option later if availability changes.

4:36 Simple example: team on Google Cloud

A simple example. A marketing team already runs everything on Google Cloud and needs a model that writes product titles in a strict format. Vertex AI supports supervised tuning for the Gemini two point five models they already use. They convert their canonical data to the Gemini format with a small script, launch a job, and compare the tuned model with their best prompt. The tuned model wins on format accuracy. They keep the canonical data and converter in Git, so they could re-run the same experiment elsewhere.

5:14 Lifecycle planning

One more consideration: lifecycle. Hosted base models are retired on a schedule, and a fine-tune of a retired base model stops working when the base does. When you pick a hosted path, note the base model's announced retirement policy, set a reminder well before any date, and budget for retraining on the successor, including re-running your full evaluation. With portable data and an automated evaluation, that migration becomes a routine task rather than an emergency.

5:47 FAQ: is custom-model inference pricier?

A question worth asking every time: does a hosted fine-tune cost more to run than the base model? It can. Some platforms price inference on custom models differently, and some require dedicated capacity to serve them. Put those numbers into your cost model before training, not after. A fine-tune that saves prompt tokens but requires expensive dedicated hosting may not save money at your volume.

6:15 Try this now

Try this now. Write a twenty-line converter from your canonical JSON lines to one other provider's format, for example from messages to Gemini's contents and parts. Run it on ten examples and validate them. Once you have one converter working, switching providers stops being scary.

6:35 Watch me do it

Watch me do it. Our canonical data is prompt and completion JSON lines. I write a short converter to the Gemini format: the system message becomes the system instruction, user turns become contents with role user and a text part, and the assistant completion becomes role model. I convert ten examples and validate them against the format described in the docs. Then I upload the full files to a storage bucket in the approved region and launch the tuning job with the Google Gen AI SDK, passing the base model, training dataset and a validation dataset. While it runs, I fill in the six-point checklist: models and regions, methods, data handling from the terms page, inference pricing for tuned models, the deprecation policy for the base, and export options, which for this closed model are none. When the job finishes, I evaluate the tuned model with the same harness and the same test set as our open-model adapter, so the comparison is fair.

7:46 Recap

Recap. Hosted fine-tuning is convenient and fast-moving. Know the current landscape, and verify it. Closed-model fine-tunes are not portable, so keep data, graders and evaluations provider-neutral. Check regions, data handling, inference pricing and lifecycle. Your next step: complete the checklist for two hosted options and one self-managed option, and write converters from your canonical format to each.