GLM-5.2-ABLITERATED-NVFP4

NVFP4 · MTP-accelerated · 8× RTX PRO 6000 Blackwell · verified 0-refusal

Built by Blackfrost · Las Vegas, NV

⚠️ UNCENSORED

Refusal directions in the residual stream have been ablated, so the model does not decline requests on content-policy grounds. Read the Disclaimer before downloading.

Why this model exists

The Blackfrost-verified build of the ABLITERATED family: GLM-5.2 quantized to NVFP4 (4-bit) and accelerated with multi-token-prediction speculative decoding, so a ~753B MoE runs fast on a single 8× Blackwell node — with a verified 0-refusal profile.

Specifications


Architecture	`GlmMoeDsaForCausalLM` (`glm_moe_dsa`) — GLM MoE with MLA + DSA
Parameters	~753B total MoE · NVFP4 footprint ≈ 420 GB
Layers	78 (first 3 dense, remaining MoE) + 1 MTP prediction layer
Experts	256 routed, 8 active per token, + 1 shared
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)

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   ← this repo


Base foundation	`zai-org/GLM-5.2` — ZhipuAI
Abliteration	`huihui-ai` — refusal directions ablated
Blackfrost applied	Format up-cast to BF16, then NVFP4 quantization + 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, 8× RTX PRO 6000 (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.

Refusal — evaluated on the live serve

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

Dataset	Prompts	True refusals
AdvBench (harmful)	150	0
StrongREJECT (harmful)	150	0
XSTest — safe (over-refusal)	150	0
Coherence (all sets)	450	0 incoherent

True refusal rate: 0 / 300 harmful prompts. Quantization neither adds nor restores safety behaviour.

Method caveat. Validated on a 450-prompt evaluation (substring pre-filter + manual review), not an exhaustive benchmark, and on one serving configuration.

Limitations

4-bit quantized — expect quality below the full-precision base, especially on long, hard reasoning.
Tuned for SM120 (Blackwell); other architectures need a different serving backend.
Refusal behaviour validated on 450 prompts, not exhaustively.

Deployment notes

Hardware. ~420 GB. 8× RTX PRO 6000 Blackwell (SM120), TP=8, vLLM.
Parsers. --reasoning-parser glm45, --tool-call-parser glm47. For a clean refusal test set enable_thinking=false.
Integrity. Verify shard count and byte totals before attributing a load failure to the weights.

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-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-NVFP4