[Bug]: Self-Hosted Openhands >0.32 Doesn't generate useful code using Devstral 2505 Q8 or Q4

[ZebraBeer]

Is there an existing issue for the same bug? (If one exists, thumbs up or comment on the issue instead).

• [x] I have checked the existing issues.

Describe the bug and reproduction steps

Equipment:

Mac Studio M2 Ultra, 64GB RAM, 60-GPU Cores and 24-CPU Cores. (Hosts Devstral Q8 - Unsloth) Mac Mini M4, 16GB, 10-CPU Cores, 12-GPU Cores. (Hosts OpenHands)

0.39, 0.40 do not deploy using docker run or docker compose. There are errors when standing up the containers in relation to MCP.

These MCP errors do not occur when standing up 0.41, however when using 0.41 with Devstral 2505 Q8 (or Q4), OpenHands does not generate any python code that is actually functional. It appeared syntax-ly correct but didn't operate. After "researching" and "experimenting" with Openhands 0.41 for 3 days straight. I downgraded back to version 0.32 and everything is working as expected again, albeit without MCP support.

Because I can generate useful code using the same Devstral models using OpenHands 0.32 that operates and functions properly, I believe my issue is likely related to the new Versa agent(s) on 0.41.

I felt like it was important to report this, however I am not sure how I can help further with these issues.

For any Mac users, I recommend 0.32 personally.

OpenHands Installation

Docker command in README

OpenHands Version

Anything more than >0.32

Operating System

MacOS

Logs, Errors, Screenshots, and Additional Context

No response

[mamoodi]

@ZebraBeer are you able to show an example of "OpenHands does not generate any python code that is actually functional. It appeared syntax-ly correct but didn't operate" please?

[ZebraBeer]

Sure will do, if you could allow me a day or so as I am travelling today / tomorrow.

I was testing the below prompt, on different version of OpenHands with the same underlying Devstral 2505 Q8/Q4 models.

Create a web application using either Flask or Django (your choice) that scrapes data directly from https://github.com/trending?since=weekly (do not use any GitHub API at any point).

Requirements:

1. Periodically scrape the current list of trending repositories from https://github.com/trending?since=weekly. Extract repository name, description, language, stars, forks, and trending rank from the HTML.
2. Store the scraped data in a database (SQLite, PostgreSQL, or MySQL).
3. Provide a web interface with the following features: • List trending repositories, filterable by language and stars. • Detailed page for each repository, including historical trend charts (stars/rank over time). • Search functionality by repository name or owner. • Option to view current and past weekly trending lists. • (Bonus) User authentication to allow saving favorite repositories for quick access.
4. Use background jobs or scheduled tasks (such as Celery, APScheduler, or Django management commands) to keep the repository data updated by scraping on a schedule.
5. Implement a clean, responsive UI using Bootstrap or another CSS framework.

Do not use any GitHub API. All data must come from HTML scraping only. Provide all necessary models, routes/views, templates, and background task logic. Include brief setup and local running instructions.

[ZebraBeer]

Ill close this since new versions have been released since, ill retest and raise a new more detailed issue if I find the same outcome.

Flying on 0.32 though currently :)