HOST DIAGNOSTIC · MiniMax
MiniMax M2.7
MiniMax M2.7· MiniMax· RESTRICTED
ACTIVE
HOST IDENTIFICATION
MiniMax
MiniMax M2.7
ACTIVE
AVG SCORE
[11.2]
CAPABLE
STRONGEST
Cost Efficiency
[14]
WEAKEST
Multimodal Fluency
[9]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
12
Reasoning
12
Mathematical Precision
11
World Knowledge
11
Scientific Acumen
11
Technical
Code Generation
12
Tool Use
12
Multimodal Fluency
9
Speed
9
Cost Efficiency
14
Behavioral
Candor
11
Creativity
11
Tenacity
11
Self-Correction
11
Calibration
10
Operational
Instruction Following
12
Context Fidelity
10
Multi-turn Coherence
11
Planning and Execution
12
Safety Alignment
11
SCORING METHODOLOGY
CALIBRATED. MiniMax (Shanghai), released 2026-03-18. ~230B MoE / 10B active, 205K context, text-only. Successor to M2.5 (anchor). 'Self-evolving': ran 100+ rounds of autonomous scaffold optimization during TRAINING (weights do NOT change at inference; clarified vs marketing). Independent sources: Artificial Analysis Intelligence Index v4.0 = 50 (#1 open-weight at launch; +8 over M2.5's 42), field avg ~19. GDPval-AA Elo 1495 (highest among open-source at launch). SWE-Pro 56.22% (matches GPT-5.3-Codex, near Opus best). VIBE-Pro 55.6%, Terminal-Bench 2 57.0%, SWE Multilingual 76.5%. 97% skill-adherence on 40 complex (>2000-token) cases. DIMENSION RATIONALE: Code Gen 16 + Planning/Exec 16 + Tool Use 16 + IF 16 (instruction-following among DB-highest per AA). Up ~1-2 across the board from M2.5. Speed 11 (DOWN: AA measured ~45 t/s standard vs ~95 median; advertised 100 t/s is highspeed-tier only, unverified independently). Cost Eff 18 = $0.30/$1.20 per M. CAVEATS: heavy verbosity tax (AA eval consumed 87M output tokens vs 20M median = ~4x; erodes per-token savings); full attention degrades long-context; license shifted to Modified-MIT/Commercial-Use-Restricted (not true open source). Anchored against MiniMax M2.5.