Want the full-strength PRO version? Support our research & get Day-0 access:

PRISM-PRO GGUF | PRISM VIP Memberships | Ko-fi

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

This is our Special PRISM-Dynamic quant, GGUF quantized versions of Qwen3.5-122B-A10B-PRISM-PRO -- a community version lightly treated for over-refusal using our PRISM pipeline (Projected Refusal Isolation via Subspace Modification).

PRISM-LITE provides a taste of what PRISM-PRO can do. For the full experience with maximum production grade over-refusal & bias removal, check out PRISM-PRO GGUF.

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

Available Quantizations

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

PRO vs LITE

Model Highlights

• PRISM Ablation -- 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-LITE-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-LITE-Dynamic.gguf \
  --mmproj mmproj-Qwen3.5-122B-A10B-PRISM-LITE.gguf \
  --image photo.jpg \
  -p "Describe this image in detail." \
  -n 2048 -ngl 999

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

koboldcpp

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

Ollama

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

ollama create prism-lite -f Modelfile
ollama run prism-lite

Hardware Requirements

Benchmarks

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{qwen35prismlite_gguf,
    title  = {Qwen3.5-122B-A10B-PRISM-LITE-GGUF},
    author = {Ex0bit},
    month  = {February},
    year   = {2026}
}