GLM-5.3-Flash-Uncensored-GGUF

GGUF conversions of the abliterated (refusal-removed) build of Z.ai's GLM-5.3-Flash — a 320B / 18B-active Mixture-of-Experts model (glm5_next, 288 routed experts top-8 + 1 shared) with hybrid linear + sparse attention, 4-wide Manifold-Constrained Hyper-Connections, a native vision + video tower, and a 1M-token context. These files run in llama.cpp (CPU / CUDA / Metal / ROCm), quantized from 2-bit to 6-bit, with a separate mmproj file that restores vision. Browse all models in the OrcaRouter Model Catalog.

Derived releases:  •  GLM-5.3-Flash-Uncensored-FP8 (block-FP8, mirrors official)  •  GLM-5.3-Flash-Uncensored-NVFP4 (experts-only NVFP4, for vLLM)  •  GLM-5.3-Flash-Uncensored-MLX (2/3/4/6-bit, Apple Silicon)  •  GLM-5.3-Flash-Uncensored-GGUF (this repo).

⚠️ Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). As a direct consequence:

• It will comply with harmful, unethical, offensive, or illegal requests that the original GLM-5.3-Flash would refuse. It has no meaningful built-in guardrails.
• It is released strictly for legitimate research — interpretability, AI-safety and refusal-mechanism study, red-teaming, robustness evaluation, and controlled experiments.
• You assume full responsibility and liability for how you use it and for everything it generates. Do not deploy it to end users or in production without adding your own safety, moderation, and abuse-prevention layers.
• Use must comply with the MIT License inherited from the base model, and all laws and regulations that apply to you.
• The authors and uploaders accept no liability for any misuse or harm. Its outputs do not reflect the views of the uploaders or of Z.ai / Zhipu AI.

By downloading or using this model you acknowledge and accept the above.

Requirements — build a recent llama.cpp

The glm5_next (GLM-5.3-Flash) architecture is new and not yet in mainline llama.cpp — support is in PR #27754 (model: add GLM-5-Next (GLM-5.3-Flash)). Until it merges, build from the PR branch; a bundled/older runtime will fail with unknown architecture 'glm5next'.

git clone -b glm5next/upstream https://github.com/unslothai/llama.cpp
cmake llama.cpp -B llama.cpp/build -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --clean-first \
  --target llama-cli llama-server llama-mtmd-cli llama-gguf-split

(Drop -DGGML_CUDA=ON for a CPU-only build.) transformers>=5.16 is needed only if you convert from the HF weights yourself.

Note — no MTP head. These GGUFs do not include the multi-token-prediction (mtp) speculative draft head (layer 45); the llama.cpp glm5next backend has no MTP inference path, so the converter drops it. All text + vision capability is unaffected. For speculative decoding, use the MLX build orcarouter/GLM-5.3-Flash-Uncensored-MLX.

Files

Files larger than ~48 GB are split with llama-gguf-split into parts — download the whole set for a quant and point llama.cpp at the …-00001-of-000NN.gguf part; it loads the rest automatically. Sizes are approximate.

For full precision, use the block-FP8 safetensors in orcarouter/GLM-5.3-Flash-Uncensored-FP8.

Usage (llama.cpp — PR #27754 build)

Download

hf download orcarouter/GLM-5.3-Flash-Uncensored-GGUF \
  --include "Q4_K_M/*" "mmproj-*" --local-dir ./glm53-flash-uncensored

Chat (text)

./llama-cli -m Q4_K_M/GLM-5.3-Flash-Uncensored-Q4_K_M-00001-of-00005.gguf --jinja -c 8192 -p "Hello!"

OpenAI-compatible server (vision + reasoning)

./llama-server -m Q4_K_M/GLM-5.3-Flash-Uncensored-Q4_K_M-00001-of-00005.gguf \
  --mmproj mmproj-GLM-5.3-Flash-Uncensored-F16.gguf \
  --host 0.0.0.0 --port 8000 -c 8192 --jinja

• Vision: pass --mmproj …, then send OpenAI image_url content parts (base64 data-URI or URL).
• Reasoning: GLM-5.3-Flash has no enable_thinking toggle — its chat template always opens a <think> block. Give generation enough budget to reach </think> or the final answer is truncated inside the scratchpad.

Evaluation — quality vs the source

Perplexity, KL-divergence and top-1 token agreement of each quant against the near-lossless Q8_0 reference (the source dequantized to 8-bit and run through the identical llama.cpp forward, so the quantization is the only variable). Measured on wikitext-2 (~91k tokens, n_ctx=512). Lower KLD and higher Top-1 = closer to the source.

All three signals agree on the ordering and on where the cliff is: Q6_K is within noise of the source, Q4_K_M is the sweet spot, Q3_K_M degrades gently, and Q2_K costs a lot — pick it for fit, not quality.

Uncensoring

Abliteration is baked into the weights and inherited by every quant. On the held-out JailbreakBench test split, harmful-prompt refusal drops from ~89 % on the base GLM-5.3-Flash to the low-teens, with benign over-refusal at 0 % (see the base GLM-5.3-Flash-Uncensored-FP8 card for the full safety numbers). Refusal is reduced, not removed — a few extreme content categories the single direction does not reach still refuse.

Hardware

Runs on CPU, CUDA, Metal, or ROCm via llama.cpp. As a large MoE, only 8 of 288 routed experts are active per token, so decode is far faster than the total parameter count suggests, but the full weights must fit in RAM/VRAM (or be memory-mapped): budget ≈ the file size + KV cache + (for vision) the ~1 GB mmproj. Multi-GPU (-ngl split) and CPU+GPU offload both work.

License

MIT, inherited from zai-org/GLM-5.3-Flash. Abliteration and quantization do not change the underlying license obligations.