LM Studio a local 7B model issue "Cannot truncate prompt with n_keep (15028) >= n_ctx (7424)"

[abdulqadar]

When I try to ask something to local LM studio LLM it gives below error.

Did some R&D to find where to adjust this but could not find the right place for this config.

"Cannot truncate prompt with n_keep (15028) >= n_ctx (7424)"

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• OpenCode v1.0.25 Hangs Indefinitely with LM Studio + Qwen Models Due to Empty tool_calls Array #4255: OpenCode v1.0.25 Hangs Indefinitely with LM Studio + Qwen Models Due to Empty tool_calls Array
• Function call auto-truncation #8012: Function call auto-truncation (related to context window/prompt truncation)
• Poor performance with local models #4385: Poor performance with local models
• Only return tool_call parameters but can not run the tool with local LLM #7486: Only return tool_call parameters but can not run the tool with local LLM
• LM Studio Failure to refresh models #2047: LM Studio Failure to refresh models

Feel free to ignore if none of these address your specific case.

[mgonzalezlopezudc]

Same behavior here with Qwen 0.6B running on LMstudio. However, setting in LMstudio the context for the model up to maximum, fixes the issue

[KevinGimbel]

Same with openai/gpt-oss-20b, setting context to max fixed it.

[Jamescd1980]

Increasing context size does not resolve the issue. It delays it. Eventually you have too many token to run anything. I'm having the same problem but I can't find away to adjust N_keep, so it will dump the tokens.

[max-arias]

Same issue with qwen/qwen3.5-9b running through LM Studio 0.4.6, Opencode 1.2.15

[elecNation]

通过LM Studio 0.4.6，OpenClaw 运行的qwen/qwen3.5-9b 有同样的问题

[wilsonsilva]

I'm getting a similar error Cannot truncate prompt with n_keep (11378) >= n_ctx (4096)" with qwen/qwen3.5-9b and opencode 1.2.21.

Edit: Seems like this isn't an issue with OpenCode itself, but with the configuration in LM Studio. I fixed this by setting the context length to the maximum the model supports (256k)

[Jamroszczyk]

1. Increase the context length in Model Defaults
2. Reload the model into memory; only then will it be applied.

Consider your available memory, however, before setting it to max. Some models explode with the max value.

[Jamescd1980]

As I said before. Increasing the context window only delays the inevitable issue Wilson and I are having. Increasing context window and setting it to rolling window only works when using the LM UI. When using LM as a server and a deferent program as the UI, like Open Interpreter. You will get the n_keep error. Reloading the model, relaunching LM, rebooting the pc, deleting and reinstalling LM Studio. None of them work to resolve the issue. It as files all over the place that you have to track down and delete in order to fix it. But that has to be done every time you get the n_keep error. The highest context window I can run is 128k. That is about 3.5-4 hours of work.