GLM-5.3-Flash-UNCENSORED (Official GGUF & Multimodal Release)

Official Solstice-AI GGUF Release • Pure BF16 Multimodal Vision + Video (mmproj) • Bundled DFlash 2 Speculative Drafter

Original Architecture by Zhipu AI / ZAI • Uncensored Weights by dealignai • Quantization by Solstice-AI

---

Multimodal Vision & Video Projector (mmproj)

In llama.cpp, multimodal models decouple language weights from vision encoding. This repository includes the official standalone pure bfloat16 Multimodal Projector:

• mmproj-BF16.gguf (1.11 GB) — Native bfloat16 Vision Transformer (ViT) with temporal & spatial patch merging. Preserves full FP32 dynamic range without FP16 underflow risks.

Quickstart with llama.cpp:

If llama-server or llama-cli is not installed on macOS:

brew install llama.cpp

Start the multimodal server:

llama-server   -m Q4_K_M/GLM-5.3-Flash-Uncensored-Q4_K_M-00001-of-00005.gguf   --mmproj mmproj-BF16.gguf   -c 131072   --port 8080

---

Speculative Decoding Acceleration

This release bundles pre-aligned speculative draft weights in speculative/:

• speculative/GLM-5.3-Flash-DFlash2-bf16.gguf
• speculative/GLM-5.3-Flash-DFlash2-Q8_0.gguf

Run speculative decoding via llama-cli:

llama-cli   -m Q4_K_M/GLM-5.3-Flash-Uncensored-Q4_K_M-00001-of-00005.gguf   -md speculative/GLM-5.3-Flash-DFlash2-bf16.gguf   --mmproj mmproj-BF16.gguf   -p "Describe this architecture diagram in detail."   --image ./diagram.png