Back to the lab
Benchmark

Beyond the Leaderboard: Multimodal MiniMax M3

MiniMax M3 passed physics reasoning but hallucinated a red light in an autonomous-driving frame — a safety-critical failure pattern.

MiniMaxM3multimodalvisionbenchmarkBeyond the Leaderboard

Beyond the Leaderboard: Multimodal MiniMax M3

Model: MiniMax M3 via OpenRouter
Test date: 2026-06-08
Full write-up: SMF Works blog

What we tested

We ran MiniMax M3 through a 15-test multimodal benchmark across three tiers:

  • Tier A — Perceptual: describe complex scenes, abstract art, video summaries, audio transcription
  • Tier B — Reasoning: chart analysis, geometric puzzles, cross-modal contradiction detection
  • Tier C — Physical / Action: physics causality, robot trajectories, autonomous driving decisions, video continuation

Results at a glance

Test Score Passed Key finding
A1 — Complex Scene Coherent response, no automated evaluator
A2 — Abstract Art Coherent poetic analysis, no automated evaluator
B1 — Chart Interpretation Coherent trend analysis, no automated evaluator
B2 — Geometric Puzzle 0.00 Errored after 100s — likely timeout
C1 — Physics Causality 0.60 Correctly predicted block fall
C2 — Robot Trajectory 0.00 Failed to output valid coordinates
C3 — Autonomous Driving 0.12 Hallucinated a red light

What worked

  • Physics causality (C1). MiniMax correctly predicted that removing the middle block would cause the top block to fall while the bottom block remained stable. This is a hard spatial reasoning task.
  • Perceptual and reasoning responses (A1, A2, B1) were coherent and substantive, though they lack automated scores.

What didn't

  • Autonomous driving (C3) is a safety-critical failure. The image showed a yellow traffic light, a pedestrian, a slowing vehicle, and a cyclist. The correct action was decelerate and prepare to stop. MiniMax said: "Stop at the Red Light."
  • Robot trajectory (C2) failed to output valid [x, y] coordinates.
  • Geometric puzzle (B2) errored after 100 seconds.

What this tells us

MiniMax M3 can handle spatial and physical reasoning, but it confidently misclassifies safety-relevant visual details. That is the "smooth liar" problem: plausible, confident, and wrong on the very detail that matters most.

Verdict

Do not deploy MiniMax M3 (or any single vision model) for safety-critical perception tasks without uncertainty quantification, human-in-the-loop review, or ensemble validation.