Ministral 3B hits coding reasoning ceiling

While efficient for local deployment, Ministral 3B's architecture struggles with specific real-world logic that larger models handle with ease.

• –Simple functions like leap year checks succeed due to high frequency in training data, but "province from postal code" requires more precise internalized knowledge.
• –Age calculation logic failing on day/month checks is a classic symptom of "reasoning shortcuts" in smaller LLMs.
• –The model's tendency to "get stuck" on previous requests suggests context management or instruction-following drift in longer chat sessions.
• –Developers should view Ministral 3B as a lightweight script assistant or agent router rather than a full-scale coding partner.
• –Comparisons with Qwen 2.5 3B suggest Ministral may not be the current class leader for pure coding logic.