MLX quantizations of Hcompany/Holo4-35B-A3B, a 35B mixture-of-experts (MoE, 3B active) vision-language model (VLM) for Computer Use, tool-driven work, and agentic workflows.
Each quantization includes the full multimodal vision tower (333 visual weights + vision_config), enabling native image and screenshot understanding in MLX-VLM, LM Studio, and Apple Silicon workflows.
Results reported by H Company from evaluations of the original Holo4-35B-A3B model across computer tasks, long workflows, and tool servers.
| Quantization | File | Size |
|---|---|---|
| 4-bit | Holo4-35B-A3B-MLX-4bit | 19.00 GB |
| 6-bit | Holo4-35B-A3B-MLX-6bit | 27.07 GB |
| 8-bit | Holo4-35B-A3B-MLX-8bit | 35.13 GB |
| Component | Architecture | Precision |
|---|---|---|
| Language Backbone | Qwen3.5 MoE (35B total, 3B active) | Quantized (4/6/8-bit affine, group_size=64) |
| Vision Tower | Vision Transformer (333 weights) | Full precision (BF16) |
| Projector | Multimodal cross-attention / MLP | Full precision (BF16) |
Vision tower weights and multimodal projectors are preserved in full precision (BF16) to ensure optimal visual comprehension, OCR, and GUI element grounding without degradation.
pip install -U mlx-vlm
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config
# Load 4-bit (or subfolder="Holo4-35B-A3B-MLX-6bit", subfolder="Holo4-35B-A3B-MLX-8bit")
model_path = "abenzerps/Holo4-35B-A3B-MLX"
subfolder = "Holo4-35B-A3B-MLX-4bit"
model, processor = load(model_path, subfolder=subfolder)
config = load_config(model_path, subfolder=subfolder)
prompt = "Describe the user interface elements shown in this screenshot."
image = ["screenshot.png"]
formatted_prompt = apply_chat_template(
processor, config, prompt, num_images=len(image)
)
output = generate(model, processor, formatted_prompt, image, verbose=True)
print(output)
# 4-bit
python -m mlx_vlm.generate \
--model abenzerps/Holo4-35B-A3B-MLX --subfolder Holo4-35B-A3B-MLX-4bit \
--image screenshot.png \
--prompt "What action should be taken next to achieve the user goal?"
# 6-bit
python -m mlx_vlm.generate \
--model abenzerps/Holo4-35B-A3B-MLX --subfolder Holo4-35B-A3B-MLX-6bit \
--image screenshot.png \
--prompt "What action should be taken next to achieve the user goal?"
# 8-bit
python -m mlx_vlm.generate \
--model abenzerps/Holo4-35B-A3B-MLX --subfolder Holo4-35B-A3B-MLX-8bit \
--image screenshot.png \
--prompt "What action should be taken next to achieve the user goal?"
Holo4-35B-A3B-MLX-4bit/model-00002-of-00004.safetensors 5.4 GBHolo4-35B-A3B-MLX-4bit/model-00003-of-00004.safetensors 5.4 GBHolo4-35B-A3B-MLX-8bit/model-00004-of-00008.safetensors 5.4 GBHolo4-35B-A3B-MLX-8bit/model-00006-of-00008.safetensors 5.4 GBHolo4-35B-A3B-MLX-8bit/model-00002-of-00008.safetensors 5.4 GBHolo4-35B-A3B-MLX-8bit/model-00003-of-00008.safetensors 5.4 GBHolo4-35B-A3B-MLX-8bit/model-00005-of-00008.safetensors 5.4 GBHolo4-35B-A3B-MLX-8bit/model-00001-of-00008.safetensors 5.3 GBHolo4-35B-A3B-MLX-4bit/model-00001-of-00004.safetensors 5.3 GBHolo4-35B-A3B-MLX-6bit/model-00005-of-00006.safetensors 5.3 GBHolo4-35B-A3B-MLX-6bit/model-00003-of-00006.safetensors 5.3 GBHolo4-35B-A3B-MLX-6bit/model-00002-of-00006.safetensors 5.3 GBHolo4-35B-A3B-MLX-6bit/model-00004-of-00006.safetensors 5.2 GBHolo4-35B-A3B-MLX-6bit/model-00001-of-00006.safetensors 5.2 GBHolo4-35B-A3B-MLX-8bit/model-00007-of-00008.safetensors 5.0 GBHolo4-35B-A3B-MLX-4bit/model-00004-of-00004.safetensors 4.4 GBHolo4-35B-A3B-MLX-6bit/model-00006-of-00006.safetensors 2.9 GBHolo4-35B-A3B-MLX-8bit/model-00008-of-00008.safetensors 540 MBHcompany/Holo4-35B-A3BHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo abenzerps/Holo4-35B-A3B-MLX ./model-folder