Latest in AI
MindLab Macaron Venti Pro on Alibaba Cloud: What to Check Before Switching
Set up the Singapore endpoint, handle reasoning and usage events separately, and test tool-call behavior before moving an existing client to Macaron Venti Pro.
Alibaba Cloud added MindLab's Macaron Venti Pro to Model Studio's international service in Singapore on November 20. The new entry is mindlab/macaron-v1.1-venti-pro, distinct from the Venti and Coding Venti variants listed earlier.
It accepts images and text, has a million-token context window, and works with OpenAI-compatible Chat Completions and Anthropic-compatible Messages clients. For an existing integration, though, the meaningful differences are in activation, stream handling and tool behavior. A successful text reply is a start, not a migration test.
Start with the Singapore workspace
The Macaron integration guide says to activate the Macaron series in Model Studio. It lists different base URL paths for each SDK, on the Singapore workspace host https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com:
- OpenAI SDK: append
/compatible-mode/v1. - Anthropic SDK: append
/apps/anthropic. For direct HTTP, the Messages route is/apps/anthropic/v1/messages.
Use a Model Studio credential, not the key from the service you are migrating away from. Create a separate provider configuration instead of replacing the existing key and endpoint globally. That keeps the old route available if a test fails.
Send a short, non-sensitive prompt before testing your full application. If the call is rejected, check the activation, workspace and model ID before debugging the prompt.
Keep reasoning separate from the answer
For the OpenAI-compatible stream, the guide distinguishes reasoning_content from answer content. It shows Python passing reasoning_effort through extra_body and Node passing it at the top level. When stream_options.include_usage is enabled, handle an event with no choices before trying to read the first choice.
Keep reasoning text out of the visible answer and avoid saving it by default in operational logs. If your existing application expects every event to contain answer text, give it a fixture with reasoning first, then answer deltas, then usage. Test that the parser neither drops the final accounting event nor feeds non-answer text into a JSON decoder.
The guide also documents preserve_thinking for passing previous reasoning to subsequent turns. Treat that as a separately evaluated setting, not something to apply when importing arbitrary history from another provider.
Test the tool behavior you rely on
The Venti Pro model card lists function calling and structured outputs, but not built-in web search. The integration guide says streamed tool-call arguments via tool_stream are unsupported. It also says n and parallel_tool_calls=false have no effect for Venti Pro.
If a workflow assumes a single action at a time, enforce that in application code. Start with a mock inventory lookup. Check that the tool name and arguments arrive in the format you expect, complete the mock call, then verify the returned answer. Test multiple requested actions without letting the model make real changes.
If you use JSON schemas, test those separately. A valid plain-text stream does not prove your parser will accept a model's structured output, especially when reasoning and output events appear together.
Budget for output as well as context
The dedicated model card lists a 1,048,576-token context window, a 131,072-token maximum output and Singapore list rates of $2 per million input tokens and $10 per million output tokens. It also lists a separate $0.10 input cache-hit rate per million tokens. The integration guide says thinking is billed at the output rate.
For illustration, 20,000 uncached input tokens and 4,000 billed output tokens would cost $0.08 at those list rates. If that same input qualifies entirely for the listed cache-hit rate, the arithmetic becomes $0.042, excluding any cache creation cost. The savings are largest on the input side; output still costs the same.
Compare actual billed usage across a few representative requests rather than assuming the quoted cache rate applies. Long documents, screenshots and tool responses can change the actual input. Higher reasoning effort can also change the amount of output you pay for.
Decide with a small side-by-side test
Run the current route and Venti Pro on the same few tasks: a short prompt, an image question, a schema response and a multi-turn tool interaction. Record time to visible answer, successful task completion and billed tokens. Keep model rollout separate from any change to prompt structure so you can tell which change mattered.
Use the Singapore route only once the region and documented provider arrangement fit your requirements. The provider listed for this model is MindLab. Check the data terms for your workspace rather than drawing a residency conclusion from the hostname alone.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams turn launches and API changes into technical writing developers can use.