the quality battery I use for every model:

1. JSON extraction
2. code generation
3. logic puzzle
4. system prompt adherence
5. hallucination resistance
6. format compliance

results across 7 models:

• Gemma 4 E4B: 6/6
• Mistral 7B: 5/6
• LFM2 24B: 4/6
• Gemma 4 E2B: 4/6
• GLM 4.7 Flash: 2/6
• Qwen3 8B: 1/6
• Llama 3.2 1B: 1/6

the biggest model scored worst.
the smallest useful model (Gemma E4B) scored best

the parameters don’t predict the quality. everyone should test before trusting

what quality tests do you run on local models? looking for ideas to add to the battery