It seems that there is no clean solution to directly store a model incl. weights into the database. I decided to store the model as h5 file in the filesystem and upload it from there into the database as a backup. For predictions I load anyway the model from the filesystem as it is much faster than getting it from the database for each prediction.

The best approach would be to save it as a file in the system and just save the path in the database. This technique is usually used to store large files like images since databases usually struggle with them.

Well this answer comes a little too late but since there are no useful answers online, based on my intensive research, I thought I'll post this insight. Maybe some will find it helpful.

I have not tried it with a keras model, but this approach worked with a pytorch model. But I can imagine that it should work for keras or tensorflow as well.

Step 1: load the model and convert to a bytes buffer

model = torch.load(r'model.pth')

buffer = io.BytesIO()
torch.save(model, buffer)
model_data = buffer.getvalue()

Step 2: create a table in your SQL Server database either in SQL directly or using ORM (I did it using Python and Flask SQL Alchemy). The column data type (where the model will be saved) should be VARBINARY(MAX) (or db.PickleType using Flask-SQL-Alchemy).

connection_string = 'DRIVER={SQL Server};Server=myServerAddress;Database=myDataBase;User ID=myUsername;Password=myPassword;Trusted_Connection=False' 
connection = pyodbc.connect(connection_string)
cursor = connection.cursor()

query="INSERT INTO Models (Model) VALUES (?)"
cursor.execute("SET IDENTITY_INSERT Models ON")
cursor.execute(query, model_data)
cursor.execute("SET IDENTITY_INSERT Models OFF")

connection.commit()

#to get the model from the db:
query="SELECT Model FROM Models"
cursor.execute(query)
row = cursor.fetchone()
model_data = row.Model

buffer = io.BytesIO(model_data)
loaded_model = torch.load(buffer)

cursor.close()
connection.close()

TA-DA

Well, you can save a model as a JSON like:

json_string = model.to_json()

And then cast it to string and save normally to the BD.

Or use native MS SQL solution for JSON fields explained here .