Blackfrost-AI/GLM-5.2-ABLITERATED-NVFP4

🔒 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 · card or Bitcoin (−10%)

---

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

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 before downloading.

---

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.

---

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.

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.