DeepSeek-R1 Runs 30 Minutes Uncapped in LM Studio

Hot take: this is usually a configuration problem, not the model “going crazy.”

• –Local reasoning models can keep producing chain-of-thought until they hit a token limit or an external stop condition, so an uncapped session can run far longer than expected.
• –The most reliable mitigation is to set a hard cap on output tokens or a wall-clock timeout in LM Studio or the serving layer.
• –Prompting for “brief reasoning” can help, but it is not a hard guarantee; the model may still ramble if the stop conditions are loose.
• –If you want something closer to a 2-minute ceiling, use a smaller context window, lower max tokens, and a stop rule in the client rather than relying on the prompt alone.