license: other
license_name: glm-5.3
license_link: LICENSE
base_model:
- zai-org/GLM-5.3
- dealignai/GLM-5.3-UNCENSORED-FP8
tags: - glm
- glm-5.3
- glm-dsa
- moe
- gguf
- uncensored
- abliterated
- maya
language: - en
- zh
- ru
pipeline_tag: text-generation
GLM-5.3 Uncensored — Maya quant (G-M24U-GPTQ)
The full GLM-5.3 (753B MoE) quantized to ~2.5 bits with GPTQ, with refusals removed, built to run on the
GLM-5.3 Uncensored on Maya engine — a fork of
Project Maya ported to the full GLM-5.3.
43–57 tokens/s on 7 GPUs (5× RTX 4090 48 GB + 2× A100 40 GB), 512K context, 2 / 40 refusals on JailbreakBench (stock: 39 / 40).
Files
| File | Size | What |
|---|---|---|
GLM-5.3-Maya-G-M24U-GPTQ-0000{1..6}-of-00006.gguf |
267 GB | the quantized model (llama.cpp glm-dsa tensor layout) |
pack/ |
2.8 GB | the Maya engine's index of the shards + dense weights + tokenizer + chat template; it finds the shards next to itself |
dir3.f32 |
5.8 MB | the refusal vector, 78 layers × 3 directions × 6144, float32 (applied at runtime) |
Run
git clone https://github.com/denisnotes357-prog/GLM-5.3-Uncensored-Maya && cd GLM-5.3-Uncensored-Maya
./setup.sh # builds the engine (decline the Flash model download it offers)
hf download nofacedfl/GLM-5.3-Uncensored-Maya --local-dir models/GLM-5.3-Uncensored-Maya
tools/glm53/run-glm53.sh # chat at http://127.0.0.1:8095, OpenAI and Anthropic APIs
The shipped config needs ~320 GB of VRAM across 7 GPUs for full speed; with less, the engine serves the experts that
do not fit from RAM (slower). The tested runtime is the Maya engine; llama.cpp compatibility of this exact file was not tested.
How it was made
- Quant (G-M24A-GPTQ). FP8 calibration statistics on 128 sequences (chat, multilingual, reasoning, code, tool
calls) → imatrix for every tensor and GPTQ for the experts' gate/up. Experts IQ2_S gate/up and IQ3_XXS down
(IQ3_S down in layers 3–6 and 74–77); attention and shared expert Q4_K, Q8_0 on the layers placed on the A100s;
NextN Q3_K/Q4_K; router and norms F32. Against FP8 on 8 held-out texts: KL 0.204, top-1 agreement 88.4%. - Donor (G-M24U).
dealignai/GLM-5.3-UNCENSORED-FP8changes onlyself_attn.o_projandmlp.shared_experts.down_proj(155 tensors); those were quantized to the same types and swapped in.
Refusals 39 → 6 of 40, speed unchanged. - Refusal vector (
dir3.f32). Three directions per layer measured on 256 harmful / 100 harmless prompts
(method of alesha-pro),
projected out of the residual stream after every layer by the engine. Refusals 6 → 2 of 40.
Full story, numbers and dead ends: HOW-IT-WAS-MADE.md.
Speed (7 GPUs, pipelined MTP window, 400-token answers)
| Text | tok/s |
|---|---|
| English | 54–57 |
| Russian | 43–45 |
| Chat at temperature 1.0 | 43–44 |
| 1M context | ~30 |
Responsible use
This model answers requests the original refuses. It is published for research, red-teaming, private and creative use
by adults. You are responsible for how you use it and for following the laws that apply to you. Outputs can be wrong
or harmful; do not rely on them for medical, legal or other high-stakes decisions.
License and credits
- Model: GLM-5.3 by Z.AI under the GLM-5.3 License (included in this repo; permissive, with a
security-review condition for Model-as-a-Service operators above US$10B revenue). - Donor tensors: dealignai/GLM-5.3-UNCENSORED-FP8, MIT.
- Engine: Project Maya by Niko1221 and the Strata contributors, MIT;
the GLM-5.3 port in GLM-5.3-Uncensored-Maya, MIT.