opencode tool_calls problem with OpenVINO/Qwen3-Coder-30B-A3B-Instruct-int4-ov

[mmoll]

Describe the bug
When using opencode against ovms with OpenVINO/Qwen3-Coder-30B-A3B-Instruct-int4-ov at some point after a few steps, I get json-output in the opencode console, I guess something with the API response of ovms is slightly wrong:

This does not happen with non-coder Qwen3 models I tested.

It does also happen with the coder model when omitting the --tool_parser qwen3coder --reasoning_parser qwen3 --enable_tool_guided_generation true options.

To Reproduce
Steps to reproduce the behavior:

1. Launch the ovms command with --pull to get the model
2. OVMS launch command: ovms --source_model OpenVINO/Qwen3-Coder-30B-A3B-Instruct-int4-ov --model_repository_path models --task text_generation --rest_bind_address 127.0.0.1 --rest_port 8000 --target_device GPU --tool_parser qwen3coder --reasoning_parser qwen3 --enable_tool_guided_generation true --cache_size 2 --log_level TRACE
3. Client command: opencode
4. See error

Expected behavior
opencode should just work?

Logs
see https://gist.github.com/mmoll/723fd621dd2f2bfb1fb5871ce4b549ab

Configuration

1. OpenVINO and ovms version: 2026.4.0

Additional context
Add any other context about the problem here.

[mzegla]

Looks like tool calls from previous conversation rounds are not properly wired back to the promt on next chat turns resulting in context slowly degradating until at some point it starts generating tool calls in wrong format. I can see you are using OVMS without python support. Could you try the same with OVMS with enabled python?
Among release assets look for package with suffix _python_on. Version with python uses different chat template rendering engine that might help here.

[mmoll]

As I'm using this on Debian/testing and there are no ubuntu26 ovms builds available I can't get a matching Python version... So I guess I have to switch to a dockerized version, or are there Ubuntu 26 builds coming soon?

[mzegla]

Ubuntu 26 is planned to be added soon, possibly in 2026.5. In such case, I'd suggest trying dockerized version for now.

[mmoll]

trying dockerized version for now.

OK, will do that.

I'd like to leave this issue open for now and I'll test again with a Ubuntu 26 build. However, if the ovms build with python_on fixes it, it would still be nice to fix the issue for python_off also.