license: mit base_model:
- Blackfrost-AI/GLM-5.2-ABLITERATED-BF16 base_model_relation: quantized pipeline_tag: text-generation library_name: vllm language: [en, zh] tags: [glm, glm-5.2, glm_moe_dsa, mixture-of-experts, moe, abliterated, uncensored, nvfp4, fp4, mtp, speculative-decoding, vllm, security-research, red-teaming]
Blackfrost
⚠️ 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.
Decode throughput on 8× RTX PRO 6000 — 2.1× faster with MTP acceleration
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-5.2source rather than from a third-party abliteration — seeBlackfrostAI/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 setenable_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.