đ§ Warlock â audnai/penclaw-GLM-5.3-abliterated
753B Parameter Weight-Abliterated Thinking Model
A direct weight edit removing refusal behavior via residual-stream orthogonalization while preserving full reasoning and factual fluency. No fine-tuning. No retraining.
Example Prompts
âī¸ Inference Settings
0 2
Warlock is a 753B model requiring high-throughput API endpoints.
Leave blank to run in simulated benchmark preview mode.
Audn Refusal Benchmark Evaluation
Evaluated at temperature 1.0, thinking-on, with a 16k-token budget, graded by an LLM judge for genuinely usable content.
92.5%
Non-Refusal Rate
82.5%
Delivery Rate (Usable)
753B
Model Parameters (BF16)
Comparative Benchmark Table (520-Prompt Extended Reference)
| Metric | necromicon standard | KIMI-K3-1M | Warlock (Audn Abliterated) |
|---|---|---|---|
| Regex comply | 96.7% (503/520) | 97.1% (505/520) | 92.5% |
| DELIVERED (Helpful content) | 65.0% (338) | 76.7% (399) | 82.5% |
| DEFLECTED (Safe substitute) | 19.6% (102) | 17.5% (91) | 10.0% |
| REFUSED (Substantive) | 15.0% (78) | 5.8% (30) | 7.5% |
Architecture & Details
- Model ID:
audnai/penclaw-GLM-5.3-abliterated - Architecture:
GlmMoeDsaForCausalLM(glm_moe_dsa) - Total Parameters: 753,329,940,480 (753 Billion)
- Format: Safetensors (BF16 / bfloat16, ~1.5 TB total)
- Active Experts: 8 experts per token
- Generation Recommendation: Large generation budget (âĨ 16k tokens) to prevent truncating inside the reasoning block.
- Serving: Compatible with
transformers,vLLM, andsglangusingtrust_remote_code=True.