You can deploy this model and start earning money today!

model-rampage/
Qwen3-0.6B-mlx-8bit

Views

No views yet

Qwen/Qwen3-0.6B (8-bit Native MLX Family Core)

This asset is an optimized execution variant of Qwen/Qwen3-0.6B, compiled natively for Apple Silicon hardware acceleration via the MLX framework under the Model Rampage Telemetry Engine.

📊 Cross-Precision Family Matrix

The comprehensive family optimization footprint table generated via our open control evaluation metrics.

Precision Target	Disk Footprint	Peak VRAM	Quantized PPL	Brain Fidelity	Active Selection
2-bit Precision	0.18 GB	0.23 GB	940136.3923	0.0%
3-bit Precision	0.25 GB	0.3 GB	74.1441	33.3%
4-bit Precision	0.32 GB	0.37 GB	29.5427	83.57%
8-bit Precision	0.6 GB	0.65 GB	24.7516	99.75%	👈 (This Repo)

Baseline Reference Model Perplexity Control: 24.6901
Standardized Evaluation Corpus Suite: Salesforce/WikiText-2 (Raw Validation Split)
Data Context Coverage Scope: Full Dataset (212 Chunks)

💻 Cross-Device Throughput Estimation Map

Projected prompt throughput estimations generated via empirical memory-bandwidth scaling configurations relative to host benchmark runtime execution.

Quant Format	🟢 M-Series Base (~100 GB/s)	🔵 M-Series Pro (~150-200 GB/s)	🟣 M-Series Max (~300-400 GB/s)	🔥 M-Series Ultra (~800 GB/s)
2-bit	104.71 tok/s	190.39 tok/s	371.26 tok/s	647.33 tok/s
3-bit	101.29 tok/s	184.16 tok/s	359.11 tok/s	626.14 tok/s
4-bit	106.18 tok/s	193.05 tok/s	376.45 tok/s	656.37 tok/s
8-bit	86.56 tok/s	157.38 tok/s	306.89 tok/s	535.09 tok/s	👈

🚀 Native Implementation Scaffolding

python

from mlx_lm import load, generate
model, tokenizer = load('model-rampage/Qwen3-0.6B-mlx-8bit')

response = generate(
    model, 
    tokenizer, 
    prompt="Your system prompt execution sequence goes here.", 
    max_tokens=128, 
    verbose=True
)

Telemetry tracked via Model Rampage Engine — Control Verification Core.