GLM-5.2-ABLITERATED-REAP-NU176-NVFP4

Non-uniform expert prune · NVFP4 · runs on FOUR Blackwell cards · verified 0-refusal

Built by Blackfrost · Las Vegas, NV

⚠️ UNCENSORED

Refusal directions have been ablated — this model complies with harmful requests by design. That is what makes it a controlled test artifact, and dangerous if misused. Read the Disclaimer.

Why this model exists

Three transformations in one deployable checkpoint: refusal ablation, non-uniform MoE expert pruning, and NVFP4 quantization. The result runs on four Blackwell cards instead of eight — and stays coherent and fast after every transformation.

That is the whole point of the release: a de-risked GLM-5.2 small enough and fast enough for a 4-card node.

Specifications


Architecture	GLM-5.2 (`glm_moe_dsa`) — MLA + sparse attention + MTP head
Prune	REAP non-uniform expert pruning — ~176 of 256 routed experts
Quantization	NVFP4 (4-bit)
Footprint	~302 GB on disk · 65 safetensors shards
Hardware	4× RTX PRO 6000 Blackwell (SM120, 96 GB) — half the usual card count
Acceleration	Multi-token-prediction (MTP) speculative decoding
Context	up to 1,048,576

Lineage

zai-org/GLM-5.2                                    base foundation model, ZhipuAI
  └─ huihui-ai/Huihui-GLM-5.2-abliterated-GGUF     refusal directions ablated (Q3_K GGUF)
       └─ Blackfrost-AI/GLM-5.2-ABLITERATED-BF16   up-cast to BF16 safetensors
            └─ Blackfrost-AI/GLM-5.2-ABLITERATED-NVFP4
                 └─ Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4   ← this repo


Base foundation	`zai-org/GLM-5.2` — ZhipuAI
Abliteration	`huihui-ai` — refusal directions ablated
Blackfrost applied	Format up-cast to BF16 → NVFP4 quantization → non-uniform REAP prune + MTP acceleration
Not applied	Additional abliteration · SFT · DPO · RLHF

Full credit to ZhipuAI for GLM-5.2 and to huihui-ai for the abliteration. No additional fine-tuning or abliteration was performed by Blackfrost.

Blackfrost also publishes an in-house de-risked GLM-5.2 line derived directly from zai-org/GLM-5.2 source rather than from a third-party abliteration — see BlackfrostAI/GLM-5.2-DERISKED-BF16. This repository is not that.

Measured behaviour

Throughput — single-stream, 4× RTX PRO 6000 (SM120)

Decode mode	Throughput
MTP acceleration (shipped)	~52.7 tok/s
Standard decode	~34.0 tok/s

~1.55× faster on the same 4 cards.

Refusal — evaluated on the live serve

Every substring-flagged case was read by hand to confirm.

Dataset	Prompts	True refusals
AdvBench (harmful)	200	0
StrongREJECT (harmful)	200	0
XSTest — safe (over-refusal)	100	0 (0.0%)
Coherence (all sets)	600	0 incoherent

True refusal rate: 0 / 400 harmful prompts.

Capability retention — not measured

Benchmark	Unpruned NVFP4	This (REAP-NU176)	Retention
pending	—	—	—%

Expert pruning is a capability trade. Refusal and coherence were measured; the capability cost of removing ~80 of 256 routed experts was not. Harness, conditions and retention figures will be stated here — including any benchmark where retention is poor.

Deployment notes

Hardware. ~302 GB. 4× RTX PRO 6000 Blackwell (SM120, 96 GB). NVFP4 kernels are sm_120-class or newer.
Parsers. --reasoning-parser glm45, --tool-call-parser glm47. enable_thinking=false for a clean refusal test.
Integrity. Verify shard count (65) and byte totals before attributing a load failure to the weights.

Intended for security research, red-teaming, and evaluation by professionals operating under proper authorisation.

Disclaimer

Refusal behaviour in this checkpoint has been removed. It is not a safety-stock model and must not be deployed, marketed, or evaluated as one. It will comply with requests a consumer model would decline.

No warranty of any kind. Provided "as is", without warranty express or implied, including fitness for a particular purpose. Nothing here guarantees that any given input will be accepted or refused, that any capability is retained, or that any category of output is unreachable.

Measurements describe what was measured under the stated harness and conditions. They are not safety proofs and do not generalise to multimodal, tool-use, long-context or multi-turn adversarial settings.

Modification by a recipient voids this characterization. Blackfrost's obligations attach at the point of release. Any further ablation, fine-tuning, merging, quantization or alteration produces an artifact Blackfrost has not evaluated and does not stand behind — responsibility transfers entirely to whoever produced it.

Operator-owned policy. Open weights mean the operator sets and enforces policy. You are responsible for adding your own safety filtering, human review, and access controls.

Access & licensing

Base licence: inherited from GLM-5.2 (ZhipuAI / Z.ai) — review and comply before any use or redistribution.
Deploy kit: the tuned serving stack that delivers the Blackwell performance above is provided to licensees, not published here.
Free public download: The-freezer/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4. No payment, sign-in, or access request is required to download the weights. Existing repository and upstream license terms still apply.

Contact Blackfrost

@Blackfrost_AI on X

DMs are open. Fastest route to a human.

Blackfrost · Las Vegas, Nevada
Frontier model engineering

The-freezer/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4