mlx-community/Muse-Spark-Open-816B-A42B-4bit

This model mlx-community/Muse-Spark-Open-816B-A42B-4bit was converted to MLX format from meta-models/Muse-Spark-Open-816B-A42B using mlx-lm version 0.32.2.

Re-uploaded on 2026-10-16 at 07:21 UTC with mlx-lm 0.32.2, which supports Muse Spark.

mlx-lm's default quantization for this model makes the routed experts of layers 1-60 4-bit and keeps the other weights in BF16. The checkpoint includes the vision encoder and takes about 533 GB, more than the memory of a single 512 GB Mac.

Use with mlx

pip install -U "mlx-lm>=0.32.2"
from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Muse-Spark-Open-816B-A42B-4bit")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)

You can also use

Downloads last month
2,433
Safetensors
Model size
822B params
Tensor type
BF16
ยท
F32
ยท
U32
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for mlx-community/Muse-Spark-Open-816B-A42B-4bit

Quantized
(25)
this model

Collection including mlx-community/Muse-Spark-Open-816B-A42B-4bit