Back to leaderboard

qwen3.5-4b-mlx

LM-STUDIO MLX

qwen3_5 · 4B · 4bit

Not Rec.

Mar 6, 2026 · Apple M4 Pro

deepseek-r1:latest

OLLAMA GGUF

qwen2 · 7.6B · Q4_K_M

Not Rec.

Jun 25, 2026 · Cortex-A78AE

Global Score
62 vs 23
Hardware Fit
35 vs 70
Quality Score
73 vs 3

Hardware

qwen3.5-4b-mlx deepseek-r1:latest
MachineMac miniNVIDIA NVIDIA Jetson Orin Nano Engineering Reference Developer Kit Super
CPUApple M4 ProCortex-A78AE
Cores146
RAM64 GB7 GB
GPUApple M4 ProIntegrated / Unknown
OSmacOS 15.7.4Ubuntu 22.04.5 LTS
Archarm64arm64
Power Modebalancedbalanced

Performance

qwen3.5-4b-mlx deepseek-r1:latest
Tokens/sec0.09.8
First chunk5548 ms1206 ms
TTFT5.5 s1.2 s
Load timeN/A17.8 s
Memory usage2.9 GB4.8 GB
Memory %5%65%

HW Fit Score Breakdown

qwen3.5-4b-mlx

Speed
0/50
TTFT
6/20
Memory
29/30

deepseek-r1:latest

Speed
32/50
TTFT
20/20
Memory
18/30

Quality

qwen3.5-4b-mlx

Reasoning
17/20
Coding
17/20
Instruction
9/20
Structured
8/15
Math
13/15
Multilingual
9/10
Reasoning: Strong Coding: Strong Instruction Following: Weak Structured Output: Adequate Math: Strong Multilingual: Strong

deepseek-r1:latest

Reasoning
0/20
Coding
1/20
Instruction
0/20
Structured
0/15
Math
2/15
Multilingual
0/10
Reasoning: Poor Coding: Poor Instruction Following: Poor Structured Output: Poor Math: Poor Multilingual: Poor

Run yours and compare

$ npm install -g metrillm@latest
$ metrillm

Requires Node 20+ and Ollama or LM Studio running

Or run without installing: npx metrillm@latest