qwen3.5:35b-a3b

qwen35moe · 36.0B · Q4_K_M

THINKING MODEL

MacBook Pro (Apple M4 Pro)

48 GB · macOS 26.3.1

Tested on March 6, 2026 · Submitted by armendarabyan

Top 67% Compare

Global Score

59 /100

Not Rec.

Hardware Fit

69/100

Quality

54/100

Get this model

🦙

Ollama

ollama pull qwen3.5:35b-a3b

View on Ollama Library

ollama.com/library/qwen3.5

Get it in LM Studio

Search and download models directly from the app

🤗

Find on HuggingFace

GGUF versions & quantizations

Hardware

Machine: MacBook Pro
CPU: Apple M4 Pro
Cores: 14 total (10 perf + 4 eff)
Frequency: 2.4 GHz
RAM: 48 GB LPDDR5
GPU: Apple M4 Pro
OS: macOS 26.3.1
Arch: arm64
Power Mode: balanced

Performance

Tokens/sec Tokens generated per second — higher is better: 30.5
Standard deviation Variation of speed across benchmark runs — lower means more consistent: ±0.2
First chunk latency Delay before the first streamed chunk arrives from runtime: 333 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: 9.8 s
Memory usage RAM consumed during inference — percentage of total system memory: 25.1 GB (52%)
Total tokens Total number of tokens generated across all benchmark prompts: 1429
Thinking tokens (est.) Estimated internal reasoning tokens — models with "thinking" generate more tokens: ~750

Score breakdown

Speed

45/50

Time to first token

0/20

Memory

24/30

Quality

Reasoning

18/20

Coding

1/20

Instruction following

7/20

Structured output

6/15

Math

13/15

Multilingual

9/10

Category levels

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

Metadata

Spec version: 0.2.1
Runtime: Ollama 0.17.7
Model format: GGUF
Hardware profile: BALANCED
Result hash: 20eaed2a347544dca74eb7e2461517faf385e053418c74954e51d950f3aab93a

Interpretation

Hardware fit: 69/100. Overall suitability: NOT RECOMMENDED (Global 59/100). Category profile: Reasoning: Strong, Coding: Poor, Instruction Following: Weak, Structured Output: Weak, Math: Strong, Multilingual: Strong.

Warnings

Significant swap activity during benchmark (+4.2 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: 17841ms for BALANCED profile)

Bench Environment

Power: Battery Swap delta: +4.2 GB CPU load: avg 13% (peak 17%)

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