qwen3.6-35b-a3b

qwen35moe · 35B · Q4_K_S

THINKING MODEL

Mac Studio (Apple M2 Max)

96 GB · macOS 26.1

Tested on April 20, 2026 · Submitted by Frank

Top 19% Compare

Global Score

82 /100

Not Rec.

Hardware Fit

80/100

Quality

83/100

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 Studio
CPU: Apple M2 Max
Cores: 12 total (8 perf + 4 eff)
Frequency: 2.4 GHz
RAM: 96 GB LPDDR5
GPU: Apple M2 Max
OS: macOS 26.1
Arch: arm64
Power Mode: balanced

Performance

Tokens/sec Tokens generated per second — higher is better: 55.2
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: 4 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: N/A
Memory usage RAM consumed during inference — percentage of total system memory: 30.3 GB (32%)
Total tokens Total number of tokens generated across all benchmark prompts: 1591
Thinking tokens (est.) Estimated internal reasoning tokens — models with "thinking" generate more tokens: ~685

Score breakdown

Speed

50/50

Time to first token

0/20

Memory

30/30

Quality

Reasoning

18/20

Coding

10/20

Instruction following

18/20

Structured output

14/15

Math

14/15

Multilingual

9/10

Category levels

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

Metadata

Spec version: 0.2.1
Runtime: LM Studio 0.4.10+1
Model format: GGUF
Hardware profile: HIGH-END
Result hash: 7c51052dab7d7455d9d21d5f6effa77b7abb48768dd161de5c5864e9d48592f6

Interpretation

Hardware fit: 80/100. Overall suitability: NOT RECOMMENDED (Global 82/100). Category profile: Reasoning: Strong, Coding: Weak, Instruction Following: Strong, Structured Output: Strong, Math: Strong, Multilingual: Strong.

Warnings

Model memory footprint is estimated via LM Studio CLI rather than measured from a fresh load.

Disqualifiers

Time to first token too high: 30000ms (maximum: 14500ms for HIGH-END profile)

Bench Environment

Power: AC CPU load: avg 10% (peak 11%)

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