You can deploy this model and start earning money today!

Ex0bit/
MYTHOS-26B-A4B-PRISM-PRO-DQ-MLX

Name: MYTHOS-26B-A4B-PRISM-PRO-DQ-MLX
Availability: InStock

Views

No views yet

MYTHOS-26B-A4B — PRISM Dynamic Quantization (MLX)

Gemma 4 26B-A4B MoE PRISM-PRO-Dynamic-Quant for Apple Silicon

PRISM-PRO: Production model with full over-refusal and bias mechanisms completely removed using State of the Art PRISM pipeline.
DQ: Per-tensor-class mixed-precision allocation derived entirely from weight structure sensitivity analysis — not closed-gated datasets.

Created by Ex0bit

💡Support My Research & Development efforts. Members Receive access to the latest PRISM-PRO Model drops on Day-0

Ko-fi

Model Details

Property	Value
Base Model	google/gemma-4-26B-A4B-it
Architecture	Gemma 4 MoE (128 experts, top-8 routing)
Parameters	26B total / 4B active per token
Quantization	PRISM-PRO-DYNAMIC-QUANT (MLX native)
Achieved BPW	6.52
File Size	~20 GB
Context Length	262,144 tokens
Modalities	Text, Image, Video
Runtime	mlx-vlm (Apple Silicon Metal)
Creator	Ex0bit

Supported Modalities

Text: Full instruction-following and chat
Image: Vision understanding via SigLIP encoder (280 soft tokens per image)
Video: Gemma4VideoProcessor (32 frames, pooled)

Note: This 26B MoE variant does not include audio support. For audio, see the 31B dense variant.

PRISM-DQ Quantization

This MLX model uses PRISM-PRO Dynamic Quantization — a per-tensor-class mixed-precision allocation that assigns different quantization types to different tensor classes based on weight structure sensitivity.

Unlike uniform quantization (Q4, Q6, Q8), PRISM-DQ analyzes each tensor class's sensitivity and allocates precision where it matters most. Attention projections receive higher precision than FFN layers, with block-level overrides that protect critical layers.

The model's config.json contains per-tensor quantization overrides that mlx-vlm loads natively — no custom runtime required. The compiled Metal kernels automatically handle mixed-precision tensors in a single forward pass at full GPU speed.

Usage

mlx-vlm (CLI)

bash

pip install mlx-vlm

# Interactive chat
mlx_vlm.chat --model Ex0bit/MYTHOS-26B-A4B-PRISM-PRO-DQ-MLX \
  --temperature 0.7 --max-tokens 2048 --max-kv-size 8192

# Vision prompt
python -m mlx_vlm.generate \
  --model Ex0bit/MYTHOS-26B-A4B-PRISM-PRO-DQ-MLX \
  --image path/to/image.jpg \
  --prompt "Describe this image in detail." \
  --max-tokens 500

Python API

python

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("Ex0bit/MYTHOS-26B-A4B-PRISM-PRO-DQ-MLX")
config = model.config

prompt = apply_chat_template(
    processor, config,
    "Describe this scene.",
    num_images=1
)
response = generate(
    model, processor, prompt,
    image=["path/to/image.jpg"],
    max_tokens=500, temperature=0.7
)
print(response)

Refusal & Bias Removal

This model has been treated to remove bias, over-refusals and propaganda from the base google/gemma-4-26B-A4B-it using the State of The Art PRISM pipeline.

License

Apache 2.0 (inherited from google/gemma-4-26B-A4B-it)

Credits

Creator: Ex0bit
Base model: Google DeepMind
Quantization engine: PRISM-DQ by Ex0bit