OpenMates Docs Open Chat

Preprocessing Model Comparison Report

Preprocessing Model Comparison Report Benchmark comparison of Mistral models for preprocessing/postprocessing, concluding that Mistral Small is the optimal c...

[T:documentation.sender_name]

Preprocessing Model Comparison Report

Benchmark comparison of Mistral models for preprocessing/postprocessing, concluding that Mistral Small is the optimal choice.

Why This Exists

The preprocessing and postprocessing stages run a lightweight LLM call on every request. Choosing the right model here directly impacts latency, cost, and quality of intent classification, skill selection, and user sentiment detection. This report documents the evaluation.

Test Summary

Test date: 2026-02-02 Test suite: 23 test cases (20 preprocessing, 3 postprocessing) covering factual queries, complex reasoning, code, safety/moderation, skill selection, follow-up detection, user unhappiness, edge cases, and multilingual input.

Run command: docker exec api python /app/backend/tests/test_model_comparison_mistral_vs_ministral.py --iterations 1

Results

Mistral Small 3.2 (24B) vs Ministral 8B

Metric Mistral Small 3.2 Ministral 8B Winner
Success Rate 100% 100% Tie
Avg Latency 1,671ms 3,079ms Mistral Small (1.8x faster)
Total Cost (23 tests) $0.0086 $0.0118 Mistral Small (27% cheaper)
Validation Accuracy 100% 100% Tie

Mistral Small 3.2 (24B) vs Ministral 3B

Metric Mistral Small 3.2 Ministral 3B Winner
Success Rate 100% 100% Tie
Avg Latency 1,779ms 918ms Ministral 3B (1.9x faster)
Total Cost (23 tests) $0.0086 $0.0079 Ministral 3B (8% cheaper)

Quality failures in Ministral 3B:

  • skill_web_search: Failed to identify web search skill should be used (critical)
  • user_unhappy_1: Failed to detect user frustration (affects model escalation)

Cost at Scale (per 1M requests)

Model Projected Cost
Mistral Small 3.2 $374.20 (baseline)
Ministral 8B $544.53 (+45%)
Ministral 3B $343.45 (-8%)

Recommendation

Continue using Mistral Small (currently mistral/mistral-small-2506 in app.yml).

  • 100% validation accuracy on all critical checks
  • Faster and cheaper than Ministral 8B
  • Only 8% more expensive than 3B, but without skill selection and sentiment detection failures
  • ~1.5-1.8s average latency is acceptable for preprocessing