README.md

---
license: mit
base_model: zandenAI/GLM-5.2-FP8-Uncensored
base_model_relation: quantized
pipeline_tag: text-generation
library_name: gguf
language:
  - en
  - zh
tags:
  - gguf
  - glm
  - glm-5.2
  - moe
  - abliterated
  - uncensored
  - quantized
  - imatrix
extra_gated_prompt: >-
  This is a quantized redistribution of a gated research artifact with reduced
  refusal behavior. By requesting access you affirm that you are of legal age in
  your jurisdiction, that you will not use it to generate CSAM, CBRN, or
  mass-harm content, and that you accept the original model license and
  disclaimer.
extra_gated_fields:
  I am of legal age in my jurisdiction: checkbox
  I will not use this model to generate illegal content: checkbox
  I accept the original license and disclaimer: checkbox
---

# GLM-5.2-Uncensored - GGUF

As a reaction to the CEO of a frontier model lobbying the media + congress
to boogeyman open source models, I release:

Imatrix GGUF quantizations of **[zandenAI/GLM-5.2-FP8-Uncensored](https://huggingface.co/zandenAI/GLM-5.2-FP8-Uncensored)**,
an abliterated (refusal-removed) build of **[zai-org/GLM-5.2-FP8](https://huggingface.co/zai-org/GLM-5.2-FP8)** -
a 754B-parameter Mixture-of-Experts model.

> All credit for the abliteration methodology and the source weights goes to
> **Zanden Kane** ([@zandenkane](https://www.threads.net/@zandenkane)). This
> repository only provides GGUF conversions for local inference.

## Method

- **Source:** FP8 -> dequantized to **BF16** -> GGUF (the BF16 source is included here).
- **Quants:** built with an **importance matrix** (imatrix) and a **dynamic, MoE-aware recipe** -
  experts at the target bit-width, with token-embeddings / output / attention kept higher
  (Q8_0 / Q6_K).
- **Calibration:** standard public imatrix calibration corpus.

## Files

| File | Type | ~Size | Notes |
|------|------|-------|-------|
| `glm52-BF16-*.gguf` | BF16 | ~1.5 TB | full-precision source; re-quant from this |
| `glm52-IQ1_M.gguf`  | IQ1_M | ~180 GB | smallest; big-RAM / Mac |
| `glm52-IQ2_M.gguf`  | IQ2_M | ~250 GB | |
| `glm52-IQ3_XXS.gguf`| IQ3_XXS | ~290 GB | |
| `glm52-IQ4_XS.gguf` | IQ4_XS | ~400 GB | best quality-per-byte at 4-bit |
| `glm52-Q4_K_M.gguf` | Q4_K_M | ~450 GB | robust 4-bit default |
| `glm52-Q6_K.gguf`   | Q6_K | ~617 GB | near-lossless |
| `glm52.imatrix`     | imatrix | small | roll your own quant levels |

*(Quants upload as they finish - some may still be in progress.)*

## Run with llama.cpp

~~~bash
./llama-server -m glm52-IQ4_XS.gguf -ngl 99 -c 16384 --host 0.0.0.0 --port 8080
~~~

For the sharded BF16, download all parts and point llama.cpp at the first shard; it auto-loads the rest.

## Hardware guide

| Quant | Fits on |
|-------|---------|
| IQ1_M / IQ2_M | 256 GB RAM box, 192-256 GB Mac, or 3-4x A100 |
| IQ3_XXS / IQ4_XS | 384-512 GB RAM, 512 GB Mac, or 6-8x A100 |
| Q4_K_M / Q6_K | 8x A100 / H100, or large-RAM CPU/offload |

## Safety

The source model removes general refusals but **preserves** categorical refusal for
CSAM / minor-exploitation content. Use responsibly and in accordance with your local
laws. The original license and disclaimer apply in full.

## License

MIT (inherited from the base model).

## Credits

- Base model: **Z.ai / zai-org** - GLM-5.2-FP8
- Abliteration + source weights: **Zanden Kane** - zandenAI/GLM-5.2-FP8-Uncensored
- GGUF conversion (imatrix + dynamic recipe): this repository