HOST DIAGNOSTIC · Google
Gemma 4 31B
Gemma 4 31B· Gemma· OPEN WEIGHTS
ACTIVE
HOST IDENTIFICATION
Google
Gemma 4 31B
ACTIVE
AVG SCORE
[14.5]
STRONG
STRONGEST
Cost Efficiency
[18]
WEAKEST
Self-Correction
[13]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
15
Reasoning
15
Mathematical Precision
16
World Knowledge
14
Scientific Acumen
15
Technical
Code Generation
14
Tool Use
14
Multimodal Fluency
14
Speed
14
Cost Efficiency
18
Behavioral
Candor
14
Creativity
14
Tenacity
13
Self-Correction
13
Calibration
14
Operational
Instruction Following
15
Context Fidelity
14
Multi-turn Coherence
14
Planning and Execution
14
Safety Alignment
15
SCORING METHODOLOGY
CALIBRATED. Google DeepMind, released 2026-04-02, Apache 2.0 open-weights. Flagship 31B Dense variant of the Gemma 4 family; built from Gemini 3 research. 256K context, 140+ languages, multimodal (text+image in, text out), configurable thinking modes, Multi-Token Prediction (speculative decoding). SOURCES: ranks #3 open model on Arena AI text leaderboard (parent family note); MMLU Pro 85.2%; AIME 2026 89.2% (strong math → Math Precision 16); 'outcompetes models 20x its size' per Google. DIMENSION RATIONALE: scored ABOVE the conservatively-rated Gemma 4 parent record on Reasoning (15 vs 14) + Math (16 vs 14) because this is the specific size carrying the Arena/AIME provenance the parent lacked. Cost Efficiency 18 + Speed 14 (runs on a workstation, not edge; the smaller E-sizes are faster/cheaper). Multimodal 14 (image in, no audio at this size). Tool Use 14 + Planning 14 (agentic-capable per family design but not a frontier agentic model). Anchored as the open-weights flagship of the Gemma 4 family. Confidence: MEDIUM-HIGH (per-size benchmarks published).