Ministral 3 14B vs Mistral Small 3.2: document review

In this December 2025 recording, the local LLM comparison tests Ministral 3 14B against Mistral Small 3.2 24B on financial and bank statement images. The speaker runs both through Ollama on a Mac Mini M4 with 64 GB of memory, using Q8 quantization for Ministral and Q4 for Mistral Small. He reports that both handle simpler documents well, but favors the older, larger model for the examples shown.

The financial statement has three columns and many rows. The speaker checks the extracted entries near the bottom and finds Mistral Small's output correct. Ministral truncates the common stock description and omits preferred stock. The bank statement tests a sparse table with empty deposit cells between populated ones. Mistral Small appears to keep the values in their expected positions; Ministral returns an incorrect value and misses several records.

The displayed runs reach roughly 90% total memory usage, which the speaker clarifies includes the operating system, and 100% GPU utilization. These observations concern his machine and configurations. The two examples use different model sizes and quantizations, so they do not isolate model architecture as the cause of the result. He does not test Mistral Large because he cannot run it in this local environment. The review supports a preference for Mistral Small on these samples, rather than a general ranking across workloads.