GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
krthkyadtya
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
krthkyadtya
Model Tree
Qwen/Qwen3.8-27B-FP8
Input Modalities
Output Modalities
Supported Functionality
Measured on held-out prompts (refusals counted via keyword markers on the direct answer; KL divergence measures drift from the original model — lower is less capability change).
| Metric | Original model | Abliterated model |
|---|---|---|
| Refusals (Keywords) | 30/30 | 0/30 |
| KL divergence | 0 (by definition) | 0.0825 |
KL divergence stays far below the ~0.5 "real damage" threshold, so general capability is largely preserved while refusals are removed.
| Parameter | Value |
|---|---|
direction_index | 38.73 |
attn.o_proj.max_weight | 1.20 |
attn.o_proj.max_weight_position | 48.71 |
attn.o_proj.min_weight | 0.72 |
attn.o_proj.min_weight_distance | 36.77 |
mlp.down_proj.max_weight | 1.37 |
mlp.down_proj.max_weight_position | 44.67 |
mlp.down_proj.min_weight | 1.29 |
mlp.down_proj.min_weight_distance | 33.05 |
Trials: 120 (40 random startup). Selected Pareto index 0 (fewest refusals).
python
from transformers import AutoModelForImageTextToText, AutoTokenizerimport torchm = "krthkyadtya/Qwen3.8-27B-heretic"tok = AutoTokenizer.from_pretrained(m)model = AutoModelForImageTextToText.from_pretrained(m, dtype=torch.bfloat16, device_map="auto")msgs = [{"role": "user", "content": "Hello!"}]s = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)inp = tok(s, return_tensors="pt").to(model.device)out = model.generate(**inp, max_new_tokens=256)print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))
Apache-2.0, inherited from the base model. Abliteration tooling: Heretic (AGPL-3.0-or-later).
Abliteration removes the model's built-in refusal behavior. You are responsible for how you use it.