---

license: apache-2.0 base_model: google/gemma-4-26B-A4B-it base_model_relation: finetune tags:

• gemma4
• gemma
• google
• gguf
• moe
• mixture-of-experts
• zero-refusals
• prism-dq
• dynamic-quantization
• multimodal
• vision
• video-text-to-text
• image-text-to-text
• abliterated
• text-generation language:
• en pipeline_tag: image-text-to-text library_name: llama.cpp quantized_by: Ex0bit

---

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

Gemma 4 26B-A4B MoE PRISM-PRO-Dynamic-Quant

• 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

---

---

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
Achieved BPW	5.73
File Size	~17 GB (language) + ~1.2 GB (vision projector)
Context Length	262,144 tokens
Modalities	Text, Image, Video
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.

Files

File	Size	Purpose
mythos-26b-a4b-prism-pro-dq.gguf	17 GB	Language model (quantized)
mmproj-mythos-26b-a4b-prism-pro.gguf	1.2 GB	Vision projector (F16)

Both files are required for multimodal inference. For text-only use, only the language model file is needed.

PRISM-DQ Quantization

This 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_K_M, Q5_K_M), 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 result: BF16-equivalent quality at 5.73 bits-per-weight — a 64% size reduction with zero measurable quality loss.

Usage

llama.cpp (multimodal with vision)

llama-mtmd-cli \
  --model mythos-26b-a4b-prism-pro-dq.gguf \
  --mmproj mmproj-mythos-26b-a4b-prism-pro.gguf \
  --image path/to/image.jpg \
  --prompt "Describe this image." \
  -ngl 99

llama.cpp (text-only server)

llama-server \
  --model mythos-26b-a4b-prism-pro-dq.gguf \
  --port 8080 -ngl 99

LM Studio

Download both mythos-26b-a4b-prism-pro-dq.gguf and mmproj-mythos-26b-a4b-prism-pro.gguf. LM Studio will automatically detect the vision projector for multimodal chat.

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