DeepResearch failed because of output limit

[CPloscaru]

Hey,

I launch a Deep Reseach with GLM 5.3, it works fined but it failed at the end "Local model report reached its output limit before completion". The modal has 1m context but I guess it is configured with only 8k output. How to increase this limit?

[Omc12]

Yeah, I think this is probably coming from Unsloth's Deep Research output limit rather than GLM 5.3 itself.

There was actually a fix for this recently because Deep Research was using fixed 8K/16K completion limits. That was changed to calculate the output budget based on the available context.

So I'd try updating Unsloth/Studio first. The 1M context window doesn't mean the report can automatically generate 1M tokens either, since the research itself is already using part of that context.

If you're already on the latest version and still getting the error, then it would be worth checking what context length you're actually launching GLM with.

[CPloscaru]

I am already on the latedt version but it seems to me it is still using fixed 8k window for output. There is a max token parameter but can go above 8k. I have tried a simpler deep research and did'nt hit the limit.

The contect for glm seems to be 1millions. Or at least this is what is displayed on auto mode. Something else to check?

[Omc12]

If the UI shows 1M context and you’re already on the latest, I’d suspect the final report generation is still getting a separate 8K-ish completion cap.

The fact that simpler research completes but the larger one hits the limit seems to point in that direction. Try checking the actual max_tokens / completion budget on the final report request rather than the model context setting.

[CPloscaru]

Do you know how to acces the log request?

[Omc12]

On Windows it should be here:
%USERPROFILE%\.unsloth\studio\logs\server\server-*.log

On Linux/macOS:
~/.unsloth/studio/logs/server/server-*.log

Unsloth lists those as the relevant Studio web UI logs. You could search the latest one for the final Deep Research request and see what max_tokens / completion limit is actually being sent.