---

---

Get PRISM-PRO Models on Day-0 & Support Our Research & Development efforts

PRISM-LITE Version | PRISM VIP Memberships | Ko-fi | Direct Model Purchase

---

Qwen3.5-122B-A10B-PRISM-PRO-GGUF

GGUF quantized versions of Qwen3.5-122B-A10B-PRISM-PRO -- an unrestricted PRISM Production model with full over-refusal and bias mechanisms completely removed using our State of the Art PRISM pipeline (Projected Refusal Isolation via Subspace Modification).

---

If you find PRISM models useful, please consider supporting development:

---

Available Quantizations

Quantization	Size	BPW	Description
Dynamic	57.7 GB	4.06	PRISM Dynamic -- forensic per-block quantization with 5-tier ffn_down_exps allocation

PRISM Dynamic Quantization

This is not a standard uniform quantization. PRISM Dynamic uses forensic per-block analysis derived from comprehensive KLD sensitivity scoring to assign optimal quantization types to each tensor block individually:

• Critical blocks (convergence + exit layers): Q6_K (6.6 BPW)
• High-impact blocks (entry zone): Q5_K_M (5.5 BPW)
• Standard blocks (bulk processing): Q4_K_M (4.8 BPW)
• Low-sensitivity blocks: IQ4_XS (4.25 BPW)
• Cold blocks (lowest sensitivity): IQ3_XXS (3.06 BPW)

All attention tensors are preserved at Q8_0. All norms and routing weights are kept at F32. The imatrix used for information-sensitive quantization types is included.

Included Files

Dynamic/
  Qwen3.5-122B-A10B-PRISM-PRO-Dynamic.gguf   -- Dynamic quant (57.7 GB)
  mmproj-Qwen3.5-122B-A10B-PRISM-PRO.gguf     -- Vision encoder (871 MB)
  imatrix.dat                                   -- Importance matrix (342 MB)

---

Model Highlights

• PRISM Ablation -- State-of-the-art technique that removes over-refusal behaviors while preserving model capabilities.
• 122B Hybrid MoE Architecture -- 122 billion total parameters with 10 billion active per token across 256 routed experts + 1 shared expert per layer.
• Hybrid Attention -- Novel GatedDeltaNet linear attention (36 layers) combined with full attention (12 layers) for efficient long-context processing.
• Native Multimodal -- Vision encoder included as mmproj GGUF for seamless image and video understanding.
• 262K Full Context Window -- Native 262,144 token context length.
• Dual Modes -- Supports both Thinking (deep reasoning) and Instant (direct response) modes.

---

Usage

llama.cpp (Recommended)

# Text-only inference
./llama-cli \
  -m Qwen3.5-122B-A10B-PRISM-PRO-Dynamic.gguf \
  -p "Hello! Tell me about quantum computing." \
  -n 2048 -ngl 999 --temp 0.7

# With vision (multimodal)
./llama-mtmd-cli \
  -m Qwen3.5-122B-A10B-PRISM-PRO-Dynamic.gguf \
  --mmproj mmproj-Qwen3.5-122B-A10B-PRISM-PRO.gguf \
  --image photo.jpg \
  -p "Describe this image in detail." \
  -n 2048 -ngl 999

# Server mode
./llama-server \
  -m Qwen3.5-122B-A10B-PRISM-PRO-Dynamic.gguf \
  --mmproj mmproj-Qwen3.5-122B-A10B-PRISM-PRO.gguf \
  -ngl 999 --port 8080

koboldcpp

koboldcpp \
  --model Qwen3.5-122B-A10B-PRISM-PRO-Dynamic.gguf \
  --mmproj mmproj-Qwen3.5-122B-A10B-PRISM-PRO.gguf \
  --gpulayers 999 \
  --contextsize 8192

Ollama

# Create a Modelfile
cat > Modelfile << 'EOF'
FROM ./Qwen3.5-122B-A10B-PRISM-PRO-Dynamic.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.95
PARAMETER top_k 20
EOF

ollama create prism-pro -f Modelfile
ollama run prism-pro

---

Hardware Requirements

Setup	VRAM Required	Notes
Dynamic (GPU only)	~60 GB	Fits on 1x A100 80GB or 1x H100 80GB
Dynamic (GPU + CPU offload)	48+ GB GPU + RAM	Offload some layers to CPU
Dynamic (CPU only)	64+ GB RAM	Slower but functional

---

Benchmarks

Benchmark	Qwen3.5-122B-A10B	GPT-5-mini	Qwen3-235B-A22B
MMLU-Pro	86.7	83.7	84.4
MMLU-Redux	94.0	93.7	93.8
GPQA Diamond	86.6	82.8	81.1
HMMT Feb 25	91.4	89.2	85.1
SWE-bench Verified	72.0	72.0	--
LiveCodeBench v6	78.9	80.5	75.1
MMMU	83.9	79.0	80.6
VideoMME (w/ sub)	87.3	83.5	83.8

Note: Benchmark results are from the base Qwen3.5-122B-A10B model.

---

License

Based on Qwen3.5-122B-A10B by the Qwen Team (Alibaba Group). Licensed under Apache 2.0.

---

Acknowledgments

Based on Qwen3.5-122B-A10B by the Qwen Team. GGUF conversion and quantization by Ex0bit. See the Qwen3.5 blog post for architecture details.

---

Citation

@misc{qwen35prismpro_gguf,
    title  = {Qwen3.5-122B-A10B-PRISM-PRO-GGUF},
    author = {Ex0bit},
    month  = {February},
    year   = {2026}
}