gemma4:12b-mlx

THINKING MODEL

MacBook Pro (Apple M1 Pro)

16 GB · macOS 26.6.2

Tested on August 24, 2026 · Submitted by AAA

Top 61% Compare

Global Score

71 /100

Not Rec.

Hardware Fit

76/100

Quality

69/100

Get this model

Get it in LM Studio

Search and download models directly from the app

🤗

Find on HuggingFace

MLX versions & conversions

Hardware

Machine: MacBook Pro
CPU: Apple M1 Pro
Cores: 10 total (8 perf + 2 eff)
Frequency: 2.4 GHz
RAM: 16 GB LPDDR5
GPU: Apple M1 Pro
OS: macOS 26.6.2
Arch: arm64
Power Mode: balanced

Performance

Tokens/sec Tokens generated per second — higher is better: 25.0
Standard deviation Variation of speed across benchmark runs — lower means more consistent: ±1.6
First chunk latency Delay before the first streamed chunk arrives from runtime: 787 ms
Time to first token How fast the model starts responding — lower is better: 30.0 s
Load time Time to load the model into memory before inference: 4.0 s
Memory usage RAM consumed during inference — percentage of total system memory: 7.1 GB (45%)
Total tokens Total number of tokens generated across all benchmark prompts: 1460
Thinking tokens (est.) Estimated internal reasoning tokens — models with "thinking" generate more tokens: ~723

Score breakdown

Speed

50/50

Time to first token

0/20

Memory

26/30

Quality

Reasoning

14/20

Coding

10/20

Instruction following

9/20

Structured output

15/15

Math

11/15

Multilingual

10/10

Category levels

Reasoning: Adequate Coding: Adequate Instruction Following: Weak Structured Output: Strong Math: Adequate Multilingual: Strong

Metadata

Spec version: 0.2.1
Runtime: Ollama 0.32.15
Model format: GGUF
Hardware profile: ENTRY
Result hash: 8f4a15e7f302d003a373519e92641b38ced64d58f3b4b4ecf84ee63adaf8956e

Interpretation

Hardware fit: 76/100. Overall suitability: NOT RECOMMENDED (Global 71/100). Category profile: Reasoning: Adequate, Coding: Adequate, Instruction Following: Weak, Structured Output: Strong, Math: Adequate, Multilingual: Strong.

Warnings

Significant swap activity during benchmark (+0.8 GB). Model may exceed available RAM — results are severely degraded.
Running on battery power — performance may be reduced.

Disqualifiers

Time to first token too high: 30000ms (maximum: 22068ms for ENTRY profile)

Bench Environment

Power: Battery Swap delta: +0.8 GB CPU load: avg 18% (peak 21%)

Run yours now

$ npm install -g metrillm@latest

$ metrillm

Requires Node 20+ and Ollama or LM Studio running

Or run without installing: npx metrillm@latest