GLM-4.7-heretic-GGUF

GLM-4.7-heretic-GGUF (RadicalNotion)

Importance-matrix quantized GGUF versions of RadicalNotionAI/GLM-4.7-heretic.

This is a strongly abliterated / decensored version of Z.ai's GLM-4.7. It achieves 0/100 refusals while retaining strong coding, reasoning, and agentic capabilities.

Note: This version has a noticeably lower refusal rate than the previous jarradh/GLM-4.7-heretic based quants (which had ~4/100 refusals), at the cost of slightly higher KL divergence.

Abliteration Details (from source model)

• Method: Heretic v1.2.0+custom (per-layer abliteration)
• Refusals: 0/100 (vs ~99/100 on the original zai-org/GLM-4.7)
• KL Divergence: 0.0748

Quantizations

Quantized with llama.cpp using importance matrices (imatrix) for better quality at lower bitrates.

Currently available:

• IQ3_M (imatrix)
• IQ4_XS (imatrix)
• Q8_0

Model Details

• Base Model: zai-org/GLM-4.7 (MIT)
• Abliterated by: RadicalNotionAI
• Architecture: glm4_moe (Mixture-of-Experts)
• Parameters: ~353–358B total / ~32B active
• Context Length: 200K+
• Thinking Support: Yes — supports interleaved/preserved thinking via <think> tags

Usage with llama.cpp

Important: You must use the --jinja flag. This enables the correct chat template that properly handles thinking blocks and tool calling.

Interactive Chat

./llama-cli \
  -m GLM-4.7-heretic.IQ4_XS.gguf \
  --jinja \
  -c 32768 \
  -ngl 99 \
  --temp 0.7 \
  --top-p 0.95

Server Mode (OpenAI-compatible API)

./llama-server \
  -m GLM-4.7-heretic.IQ4_XS.gguf \
  --jinja \
  --port 8080 \
  -ngl 99 \
  -c 32768

Tips for Thinking Mode:

• The model thinks by default.
• To see the internal <think> reasoning, you can add the --special flag in some setups.
• In compatible servers you can control thinking with chat_template_kwargs (e.g. enable_thinking).

Links

• Source abliterated model: RadicalNotionAI/GLM-4.7-heretic
• Previous (milder) quants: TanyaaDL/GLM-4.7-heretic-GGUF-v1

Disclaimer

This is an uncensored model. Safety alignments have been removed via abliteration. It may generate content that some users find offensive or inappropriate. Use responsibly and at your own risk.