qwen3.5-9b-mlx

qwen3_5 · 9B · 4bit

THINKING MODEL

Mac mini (Apple M4)

24 GB · macOS 26.3

Tested on March 11, 2026

Top 71% Compare

Global Score

58 /100

Marginal

Hardware Fit

68/100

Quality

54/100

Get this model

Get it in LM Studio

Search and download models directly from the app

🤗

Find on HuggingFace

MLX versions & conversions

Hardware

Machine: Mac mini
CPU: Apple M4
Cores: 10 total (4 perf + 6 eff)
Frequency: 2.4 GHz
RAM: 24 GB LPDDR5
GPU: Apple M4
OS: macOS 26.3
Arch: arm64
Power Mode: balanced

Performance

Tokens/sec Tokens generated per second — higher is better: 16.7
Standard deviation Variation of speed across benchmark runs — lower means more consistent: ±0.7
First chunk latency Delay before the first streamed chunk arrives from runtime: 23 ms
Time to first token How fast the model starts responding — lower is better: 14.1 s
Load time Time to load the model into memory before inference: N/A
Memory usage RAM consumed during inference — percentage of total system memory: 7.8 GB (33%)
Total tokens Total number of tokens generated across all benchmark prompts: 1617
Thinking tokens (est.) Estimated internal reasoning tokens — models with "thinking" generate more tokens: ~614

Score breakdown

Speed

36/50

Time to first token

3/20

Memory

29/30

Quality

Reasoning

10/20

Coding

16/20

Instruction following

4/20

Structured output

9/15

Math

10/15

Multilingual

5/10

Category levels

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

Metadata

Spec version: 0.2.1
Runtime: LM Studio 0.4.6+1
Model format: MLX
Hardware profile: ENTRY
Result hash: 0e0e2a1ff35248290d303ba7311e82c4c6431c38ddb14fb311c4ffe8fc39e609

Interpretation

Hardware fit: 68/100. Overall suitability: MARGINAL (Global 58/100). Category profile: Reasoning: Adequate, Coding: Strong, Instruction Following: Poor, Structured Output: Adequate, Math: Adequate, Multilingual: Adequate.

Warnings

Token throughput is estimated from LM Studio output because native token stats were unavailable. Compare tok/s across backends cautiously.
Model memory footprint is estimated via LM Studio CLI rather than measured from a fresh load.

Bench Environment

Power: AC CPU load: avg 25% (peak 34%)

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