From 0a1bd25869c122f5bc2de51d4db6dcc8a69f7d4b Mon Sep 17 00:00:00 2001 From: Rodrigo Rodriguez Date: Sun, 15 Feb 2026 20:27:35 +0000 Subject: [PATCH] fix: Increase default n_predict to 512 for DeepSeek R1 reasoning DeepSeek R1 model outputs reasoning_content first, then content. With n_predict=50, responses were truncated during reasoning phase. Increased to 512 to allow full reasoning + response. Co-Authored-By: Claude Sonnet 4.5 --- src/llm/local.rs | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/llm/local.rs b/src/llm/local.rs index b96c1773a..6f6127bb2 100644 --- a/src/llm/local.rs +++ b/src/llm/local.rs @@ -386,8 +386,8 @@ pub fn start_llm_server( let n_predict = config_manager .get_config(&default_bot_id, "llm-server-n-predict", None) - .unwrap_or_else(|_| "50".to_string()); - let n_predict = if n_predict.is_empty() { "50".to_string() } else { n_predict }; + .unwrap_or_else(|_| "512".to_string()); // Increased default for DeepSeek R1 reasoning + let n_predict = if n_predict.is_empty() { "512".to_string() } else { n_predict }; let n_ctx_size = config_manager .get_config(&default_bot_id, "llm-server-ctx-size", None)