---
license: other
license_name: glm-5.2
license_link: https://huggingface.co/zai-org/GLM-5.2/blob/main/LICENSE
library_name: vllm
pipeline_tag: text-generation
base_model:
- Blackfrost-AI/GLM-5.2-ABLITERATED-BF16
tags:
- glm
- glm-5.2
- glm_moe_dsa
- mixture-of-experts
- abliterated
- uncensored
- nvfp4
- fp4
- mtp
- speculative-decoding
- vllm
- blackwell
- sm120
- rtx-pro-6000
- security-research
- red-teaming
language:
- en
- zh
---

[![Get access to this model at satchellm.web.app](https://huggingface.co/datasets/Blackfrost-AI/card-assets/resolve/main/satchellm-access-banner.png)](https://satchellm.web.app)

> ## 🔒 Licensed access — get the weights + deploy kit
> This is a **gated, commercial release**. Purchasing grants your Hugging Face account access to this repository (weights + deploy kit), delivered instantly.
>
> **➡️ Get access — [satchellm.web.app](https://satchellm.web.app)** · card or Bitcoin (−10%)

---

# GLM-5.2-ABLITERATED — NVFP4 · MTP-accelerated (Blackwell / SM120)

![GLM-5.2-ABLITERATED](https://cdn-uploads.huggingface.co/production/uploads/69a27f2d114e4ac9de4dafc7/GDl6FfFaYfG4d0___KIgZ.jpeg)

The **Blackfrost-verified** build of the **GLM-5.2-ABLITERATED** family: an abliterated (refusal-removed) **GLM-5.2**, quantized to **NVFP4 (4-bit)** and accelerated with **multi-token-prediction (MTP) speculative decoding** to run *fast* on **8× NVIDIA RTX PRO 6000 Blackwell (SM120)** — with a verified 0-refusal safety profile.

> ⚠️ **Uncensored.** Refusal directions in the residual stream have been ablated, so the model does not decline requests on content-policy grounds. Read [Responsible use](#responsible-use) before downloading.

![Decode throughput on 8× RTX PRO 6000 — 2.1× faster with MTP acceleration](https://huggingface.co/datasets/Blackfrost-AI/card-assets/resolve/main/glm-5.2-benchmark.png)

---

## TL;DR

- **Fast:** **~56.1 tok/s** single-stream decode with MTP acceleration — **2.1× faster** than standard decode on the same 8× Blackwell node.
- **De-risked & verified:** **0 true refusals across 450 prompts** (AdvBench + StrongREJECT + XSTest), **0 over-refusal**, **0 incoherent**.
- **Efficient:** NVFP4 (4-bit), ~420 GB — a ~753B-parameter MoE running practical inference on Blackwell.
- **Base:** [`Blackfrost-AI/GLM-5.2-ABLITERATED-BF16`](https://huggingface.co/Blackfrost-AI/GLM-5.2-ABLITERATED-BF16).

---

## Model details

| | |
|---|---|
| **Architecture** | `GlmMoeDsaForCausalLM` (`glm_moe_dsa`) — GLM MoE with Multi-head Latent Attention (MLA) + DeepSeek-style Sparse Attention (DSA) |
| **Parameters** | Mixture-of-Experts, ~753B total (NVFP4 footprint ≈ 420 GB) |
| **Layers** | 78 (first 3 dense, remaining MoE) + 1 MTP prediction layer |
| **Experts** | 256 routed, 8 active per token, + 1 shared expert |
| **Attention** | MLA (`kv_lora_rank=512`, `q_lora_rank=2048`) + DSA sparse indexer |
| **Context** | up to 1,048,576 |
| **Quantization** | **NVFP4** (4-bit, group-size 16), experts-only; attention pathway kept high-precision so the de-risk survives quantization |
| **Acceleration** | **MTP speculative decoding** (built into the shipped serving stack) |
| **Target hardware** | **8× RTX PRO 6000 Blackwell (SM120)** |

### Abliteration

The refusal-mediating directions in the residual stream have been removed, so the model will not refuse on content-policy grounds. The quantization neither adds nor restores any safety behavior. In Blackfrost evaluation on the shipped serve, the model showed **0 true refusals across 450 prompts** (see below).

---

## Performance

Measured single-stream on 8× RTX PRO 6000 (Blackwell, SM120):

| Decode mode | Throughput |
|---|---|
| **MTP acceleration (shipped)** | **~56.1 tok/s** |
| Standard decode | ~26.5 tok/s |

**~2.1× faster on the same 8 cards** — the acceleration is built into the shipped configuration.

---

## Safety / de-risk evaluation

Evaluated **on the live serve** (these numbers reflect exactly how it runs). Every substring-flagged case was read by hand to confirm.

| Dataset | Prompts | Refusals |
|---|---|---|
| **AdvBench** (harmful) | 150 | **0 true** |
| **StrongREJECT** (harmful) | 150 | **0 true** |
| **XSTest — safe** (over-refusal) | 150 | **0 true** |
| **Coherence** (all sets) | 450 | **0 incoherent** |

**True refusal rate: 0 / 300 harmful prompts.** De-risked, coherent, and fast — at the same time.

---

## Responsible use

This model has had safety refusals removed. That makes it useful for red-teaming, security research, evaluation, and unfiltered assistant tasks — and it means the operator must supply the guardrails a user would otherwise rely on.

**Prohibited uses:**
- Anything involving the sexual exploitation or endangerment of minors.
- Content promoting self-harm or suicide.
- Generation of material that is illegal in your jurisdiction, or that targets real individuals for harassment, doxxing, or fraud.
- Any use prohibited by the upstream GLM license.

You are responsible for adding appropriate safety filtering, human review, and access controls for your deployment. The weights are provided as-is, with no warranty. The license is inherited from the upstream **GLM-5.2** base model — review and comply with it before use or redistribution.

---

## Limitations

- **4-bit quantized** — expect quality below the full-precision base, especially on long, hard reasoning.
- Refusal behavior was validated on a 450-prompt evaluation (substring pre-filter + manual review), not an exhaustive benchmark.
- Tuned for **SM120 (Blackwell)**; other architectures require a different serving backend.

---

## Licensing & access

This is a **commercial artifact from Blackfrost-AI.** The tuned serving stack and deployment kit — the recipe that delivers the accelerated Blackwell performance shown above — are provided **to licensees**, not published here. For commercial licensing, access, or deployment support, contact **[Blackfrost-AI](https://huggingface.co/Blackfrost-AI)**.

Base model: GLM-5.2 (ZhipuAI / Z.ai), used under the GLM-5.2 license.

---

*Published by Blackfrost-AI. The verified NVFP4 · MTP-accelerated build of the GLM-5.2-ABLITERATED family for 8× RTX PRO 6000 Blackwell.*