* llama-bench : size the batch to the prompt for models without memory Assisted-by: Claude Opus 5.5 * server : read up to 16 lfm3-d1 questions in one pass Assisted-by: Claude Opus 5.5