abenzerps/Spark-X2.6-4B-MLX-4bit

Spark-X2.6-4B MLX 4-bit

MLX 4-bit quantization of XHToken/Spark-X2.6-4B, an on-device hybrid-attention language model (4B parameters) engineered for high-efficiency agent execution. Native context: 1,048,576 tokens (1M).

Benchmarks

Benchmark results reported by XHToken for Spark-X2.6-4B in thinking mode.

MLX files

Quantization	Repository	Size
8-bit	Spark-X2.6-4B-MLX-8bit	4.37 GB
6-bit	Spark-X2.6-4B-MLX-6bit	3.34 GB
4-bit	Spark-X2.6-4B-MLX-4bit	2.32 GB

Includes upstream chat_template.jinja. Checksums: SHA256SUMS.txt.

Usage

MLX-LM CLI

# Install MLX-LM
pip install mlx-lm

# Text generation
python -m mlx_lm generate \
  --model abenzerps/Spark-X2.6-4B-MLX-4bit \
  --prompt "Hello! Introduce yourself in one sentence." \
  --max-tokens 512

Python API

from mlx_lm import load, generate

model_id = "abenzerps/Spark-X2.6-4B-MLX-4bit"
model, tokenizer = load(model_id)

response = generate(
    model,
    tokenizer,
    prompt="Hello! Introduce yourself in one sentence.",
    max_tokens=512,
    verbose=True,
)
print(response)

Source

• Source model: XHToken/Spark-X2.6-4B
• Source license: Apache-2.0