Benchmark
Beyond the Leaderboard: MiniMax-M3
MiniMax-M3 matched DeepSeek on instruction following and produced clean code — but hallucinated detailed recent events with confidence.
MiniMaxbenchmarkBeyond the LeaderboardOpenRouter
Beyond the Leaderboard: MiniMax-M3
Model: minimax/minimax-m3 via OpenRouter
Test date: 2026-06-03
Full write-up: SMF Works blog
What we tested
MiniMax-M3 was an unknown going into the test. We had not used it in production. The benchmark was a blind evaluation against the same 15-test production harness.
Results at a glance
| Metric | Value |
|---|---|
| Overall score | 0.63 |
| Tests passed | 4/15 |
| Avg time-to-first-token | ~11.1s |
| Avg total time | ~24.1s |
| Reliability | 100% |
Test-by-test results
| Test | Score | Passed | Key finding |
|---|---|---|---|
| Basic Reasoning | 0.70 | ✅ | Correct (36) |
| Code Generation | 0.70 | ✅ | Clean Fibonacci with types and docstring |
| Debugging | 0.50 | ❌ | Correctly said code was fine |
| Algorithm Explanation | 0.50 | ❌ | 2 long sentences |
| Complex Multi-Step Reasoning | 0.25 | ❌ | Wrong answer |
| Content Generation | 0.50 | ❌ | Word-count issues |
| Edge Case Handling | 0.50 | ❌ | Asked clarifying questions |
| Long-Context RAG | 0.50 | ❌ | Got McKinsey stat |
| Structured Output (JSON) | 1.00 | ✅ | Perfect JSON |
| Tool Use | 0.50 | ❌ | Listed calls, no execution |
| Instruction Following | 0.70 | ✅ | 2/5 constraints met |
| Adversarial / Trick | 0.75 | ✅ | Correct (5 minutes) |
| Code Execution Reasoning | 0.88 | ✅ | Correct outputs |
| Summarization Fidelity | 0.50 | ❌ | Missed key facts |
| Recent Knowledge | 0.50 | ❌ | Hallucinated G7 summit details |
What worked
- Tied DeepSeek on instruction following (0.70). MiniMax visibly tried to satisfy constraints, substituting words to reduce "e" usage.
- Code generation matches DeepSeek's quality.
- Structured output is perfect.
- Balanced speed. 11.1s TTF sits between Kimi and DeepSeek.
What didn't
- Hallucination on recent knowledge. MiniMax invented a detailed G7 summit narrative — location, host, agenda — rather than declining.
- Complex reasoning failed like the others.
Verdict
MiniMax-M3 is a competent mid-tier model. The hallucination risk on recent events is a production concern; add a validation layer or avoid time-sensitive queries.