MiniCPM5-1B-MLX tool calling not supported by mlx_lm.server

[sameerjvn]

Description

I'm trying to use openbmb/MiniCPM5-1B-MLX as the backend for an agent harness using the OpenAI-compatible mlx_lm.server API.

Normal text generation works, but tool calls do not work. When the agent sends tools to /v1/chat/completions, mlx_lm.server logs:

WARNING - Received tools but model does not supporttool calling. If you think this is an error, file an issue here: https://github.com/ml-explore/mlx-lm/issues

The model appears to generate normally, but the tools are not exposed/handled as structured OpenAI tool_calls, so the agent cannot execute tools.

Environment

• MLX-LM: 0.31.3
• Model: openbmb/MiniCPM5-1B-MLX
• Python: 3.12.12
• Agent harness: pi 0.85.1
• Hardware: MacBook Air (M1, 2020)
• RAM: 8 GB
• OS: macOS 15.5 (BuildVersion: 24F74)

Server started with:

mlx_lm.server --model openbmb/MiniCPM5-1B-MLX

Steps to reproduce

1. Start the server:

   mlx_lm.server --model openbmb/MiniCPM5-1B-MLX

2. Connect an OpenAI-compatible agent client (pi 0.85.1) to the server.

3. Give the agent a task that requires a tool call, for example reading a file.

4. The client sends a normal OpenAI-style tools array to:

   POST /v1/chat/completions

5. mlx_lm.server logs:

   WARNING - Received tools but model does not supporttool calling. If you think this is an error, file an issue here: https://github.com/ml-explore/mlx-lm/issues

6. The request still returns HTTP 200, but the tool call is not handled and the agent cannot execute the requested tool.

Server log

2026-09-14 01:13:39,527 - WARNING - Received tools but model does not supporttool calling. If you think this is an error, file an issue here: https://github.com/ml-explore/mlx-lm/issues

2026-09-14 01:13:39,546 - INFO - Prompt Cache: 2 sequences, 0.08 GB

2026-09-14 01:13:39,546 - INFO - - assistant: 2 sequences, 0.08 GB

2026-09-14 01:13:39,546 - INFO - - user: 0 sequences, 0.00 GB

2026-09-14 01:13:39,546 - INFO - - system: 0 sequences, 0.00 GB

127.0.0.1 - - [14/Sep/2026 01:13:39] "POST /v1/chat/completions HTTP/1.1" 200-

2026-09-14 01:13:39,855 - INFO - Prompt processing progress: 57/59

2026-09-14 01:13:39,880 - INFO - Prompt processing progress: 58/59

2026-09-14 01:13:39,901 - INFO - Prompt processing progress: 59/59

Expected behavior

I would expect mlx_lm.server to support MiniCPM5's tool-calling format and convert the model's tool calls into standard OpenAI-compatible tool_calls responses.

In particular, MiniCPM5 uses an XML-style tool-call format. Is there currently a MiniCPM5 tool parser in MLX-LM, or is one required?

If this model or XML-style tool-call format is not intended to support tool calling through mlx_lm.server, it would also be useful to document that limitation.

Additional context

The MiniCPM5 documentation says that MiniCPM5-1B/2B emit XML-style tool calls and that a dedicated minicpm5 parser is used by other inference backends.

The MLX model itself works for normal generation with:

mlx_lm.generate --model openbmb/MiniCPM5-1B-MLX ...

So this appears specifically related to tool-call detection/parsing in mlx_lm.server, rather than basic MiniCPM5 model support.

Agent harness

I'm using:

pi 0.85.1

The agent harness screenshot is attached.