deepseek-r1:14b

qwen2 · 14.8B · Q4_K_M

THINKING MODEL

MacBook Air (Apple M4)

32 GB · macOS 26.3

Tested on March 2, 2026

Top 99% Compare

Global Score

26 /100

Not Rec.

Hardware Fit

50/100

Quality

10/100

Get this model

🦙

Ollama

ollama pull deepseek-r1:14b

View on Ollama Library

ollama.com/library/deepseek-r1

🤗

Find on HuggingFace

GGUF versions & quantizations

Hardware

Machine: MacBook Air
CPU: Apple M4
Cores: 10 total (4 perf + 6 eff)
Frequency: 2.4 GHz
RAM: 32 GB LPDDR5
GPU: Apple M4
OS: macOS 26.3
Arch: arm64
Power Mode: balanced

Performance

Tokens/sec Tokens generated per second — higher is better: 11.4
Standard deviation Variation of speed across benchmark runs — lower means more consistent: ±0.1
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: 1.0 s
Memory usage RAM consumed during inference — percentage of total system memory: 16.5 GB (52%)
Total tokens Total number of tokens generated across all benchmark prompts: 1394
Thinking tokens (est.) Estimated internal reasoning tokens — models with "thinking" generate more tokens: ~1038

Score breakdown

Speed

20/40

Time to first token

0/30

Memory

30/30

Quality

Reasoning

1/20

Coding

3/20

Instruction following

1/20

Structured output

0/15

Math

0/15

Multilingual

5/10

Category levels

Reasoning: Poor Coding: Poor Instruction Following: Poor Structured Output: Poor Math: Poor Multilingual: Adequate

Metadata

Spec version: 0.2.0
Runtime: Ollama 0.17.4
Model format: GGUF
Hardware profile: BALANCED
Result hash: c78fb619362ffaaf8b15abad81fd394ec172f6c1806361b5e97384876d10018a

Interpretation

Hardware fit: 50/100. Overall suitability: NOT RECOMMENDED (Global 26/100). Category profile: Reasoning: Poor, Coding: Poor, Instruction Following: Poor, Structured Output: Poor, Math: Poor, Multilingual: Adequate. Warning: model produced very low accuracy on quality tasks — results may be unusable despite good hardware performance.

Warnings

Model produced very low accuracy on quality tasks — results may be unusable despite good hardware performance.

Disqualifiers

Time to first token too high: 30000ms (maximum: 20705ms for BALANCED profile)

Run yours now

npx metrillm@latest bench

Requires Node 20+ and Ollama running