IBM Cloud Drops watsonx.ai Fine Tuning for OpenShift AI

If your team tuned models in watsonx.ai, the tuning button is gone. IBM's release notes say watsonx.ai fine tuning was removed on 10 September 2026, with Red Hat OpenShift AI on IBM Cloud as the stated replacement. Three more removals sit in the same notes, and two of them land this week.

What IBM removed in watsonx.ai fine tuning

The watsonx release notes list "Removal of fine tuning" under 10 September 2026. Fine tuning of foundation models is removed, including the LoRA and QLoRA APIs, the Fine Tuning SDK and Tuning Studio. IBM posted the deprecation notice on 7 August. The watsonx.ai Python SDK docs add that the feature is also being removed from IBM Cloud Pak for Data starting with version 6.0.0.

What the notes do not say: what happens to tuned models you already deployed, or to old experiments. I found nothing either way. Ask IBM support before you assume a deployed tuned model keeps serving.

Export your tuning assets before anything else

According to the SDK docs, on SaaS the experiment results are stored in your own Cloud Object Storage. That means your training data and run logs sit in a bucket you control. Pull them down with the IBM Cloud CLI:

ibmcloud login --sso
ibmcloud plugin install cloud-object-storage

# get the COS instance ID (keep only the part after ::), then set it
ibmcloud resource service-instance "<cos-instance-name>" --id
ibmcloud cos config crn

# find the files, then download them
ibmcloud cos objects --bucket <project-bucket> --output json
ibmcloud cos object-get --bucket <project-bucket> --key <path/to/train.jsonl> ./export/train.jsonl

This lists the bucket attached to your project and downloads one object per call, so script the second command for the files you need. Grab the training JSONL, the run logs and any evaluation outputs. Write down the base model and hyperparameters of each run, because you will need them to reproduce the result elsewhere.

One limit: the SDK docs describe fetching data, logs and the tuned model ID. They do not describe exporting tuned model weights, so do not plan around that.

Rebuild the job on Red Hat OpenShift AI

Red Hat's route is Training Hub, which it describes as its unified fine-tuning interface. In the Red Hat material I read, it covers SFT, OSFT, LoRA and GRPO, on a Kubeflow Trainer backend or, from OpenShift AI 3.5, a Ray backend. I did not see QLoRA listed, so if you depended on it, verify before you commit.

Red Hat's LoRA on Ray walkthrough needs OpenShift AI 3.5 or later with the dashboard, workbenches and Ray components enabled. It also needs ReadWriteMany storage (50 GiB recommended) and an NVIDIA GPU worker, with A100 or H100 recommended. The core call looks like this:

from training_hub import lora_sft

lora_sft(
    model_path="<base-model-path>",
    data_path="<pvc-mount>/train_data.jsonl",
    ckpt_output_dir="<pvc-mount>/checkpoints/lora",
    num_epochs=1,
    max_seq_len=512,
    learning_rate=1e-4,
    lora_r=16,
    lora_alpha=32,
)

Those values come from Red Hat's small demo run, not a recommendation for your data. Swap in the hyperparameters you saved from your old watsonx.ai runs. The adapter lands on the shared volume, and Red Hat's evaluation step loads it with PeftModel.from_pretrained and then merge_and_unload(). The article does not cover serving the adapter, so budget time for that part.

The real change is who runs the machinery. You used to call an API. Now you own GPU nodes, storage and job scheduling on a ROKS cluster.

Three more watsonx.ai removals to check

• 8 October, already done: custom foundation models and deploy on demand foundation models are no longer available in the AWS Mumbai region, which IBM attributes to updated infrastructure and service availability. No migration link is given. Notebook sharing by URL was also removed that day.
• 13 October: the Decision Optimization runtime do_20.1. New assets could not use it after 10 September, and assets still on it become inaccessible after the 13th. IBM says to migrate to do_22.2 (CPLEX 22.2). Affected regions are Dallas, Frankfurt, Tokyo, London, Sydney, Toronto and Mumbai (AWS).
• 15 October: unstructured synthetic data is removed from the Synthetic Data Generator. After that, only structured synthetic data is available.

What to do on Monday morning

1. Export training data, logs and hyperparameters from your Cloud Object Storage bucket today.
2. Search your workspaces for do_20.1 runtimes and unstructured synthetic data jobs, and switch them before the 13th and 15th.
3. Check whether any custom or deploy on demand models run in AWS Mumbai.
4. Ask IBM about your deployed tuned models.
5. Run one LoRA job on a single GPU in OpenShift AI and score it against your old tuned model with the same eval set.

Managed AI features come with an expiry date, so keep your training data, hyperparameters and evals in storage you own and the next removal notice becomes a calendar entry instead of an incident.