A catalog of AI models in Microsoft Foundry that you can discover, compare, and deploy using Azure’s built‑in tools for evaluation, fine‑tuning, and inference
Model on HuggingFace: https://huggingface.co/Qwen/Qwen3.8-27B
AFAIK this is the official Qwen 3.8 27B.
In fact we want to use a fine-tuned variant of it because the official one is in the Foundry catalog anyways and can be used from there without a custom upload. But the official one seemed a good starting point for testing and debugging because it's a popular one that should be well supported...
Check model registration:
azd ai models show --name qwen-3-8-27b --version 4
Shows the custom/uploaded model correctly, same in Foundry AI Portal.
Verify the model files:
All model files are present. I only had to remove the "crc32.txt" from the HuggingFace package or else it complained about that file with the same error on deployment. The "manifest.base.json" is definitely not in the source folder, I have no idea where it is coming from, apparently something injects it somewhere and complains about it afterwards.
Check the base model: I've set "FW-Qwen3.8-27B" as base model which seems correct to me.
Verify the upload workflow: I've used the CLI command from in Azure AI Portal, the source parameter is present and points to the local model folder.
If all these checks pass and the issue persists, I would recommend raising a Microsoft support request
We already created a support ticket for this, answered all AI questions in the ticket responses and it then created this thread. Please don't just refer me back to AI.
This is the CLI upload command we used:
azd ai models create \
--project-endpoint "https://xxxxxxxxxxx.services.ai.azure.com/api/projects/xxxxxxx" \
-s "<REDACTED>" \
--name "qwen3-8-27b" \
--version "4" \
--base-model "azureml://registries/azureml-fireworks/models/FW-Qwen3.8-27B/versions/1" \
--weight-type FullWeight \
--publisher "Fireworks" \
--source /mnt/temp/qwen \
--debug
And here is the error in the Foundry AI Portal, as you can see we've been able to deploy multiple official models from the catalog, but not a single custom one so far: