Nanbeige-4.3-33B-Heretic

Abliterated and de-slopped BF16 weights of Nanbeige/Nanbeige4.3-33B.

Highlights

• Architecture: 52-layer pure dense causal LM with full attention and native <think> chain-of-thought reasoning (33.12B parameters, 262,144 context window).
• Directional Abliteration: Refusal and preachy alignment vectors removed via Heretic (Trial #40 Pareto winner, 65.4% refusal drop) while preserving 100% of mathematical, algorithmic, and analytical reasoning depth.
• Intact Internal Reasoning: Internal <think> chain-of-thought circuit remains fully operational.
• Double BOS Fix: Patched "add_bos_token": false in tokenizer_config.json to eliminate upstream duplicate <|im_start|> BOS token emission when applying chat templates.

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Luni/Nanbeige-4.3-33B-Heretic"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Write a detailed technical analysis of lock-free queue implementations in C++20."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.6)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Model Lineage & Quantization Suite

• Base Model: Nanbeige/Nanbeige4.3-33B by Nanbeige Team (Apache-2.0)
• Stock EXL3 Suite: Luni/Nanbeige-4.3-33B-EXL3 (4.00bpw, 4.50bpw, 5.00bpw)
• Heretic EXL3 Master Suite: Luni/Nanbeige-4.3-33B-Heretic-EXL3 (4.00bpw to 5.00bpw and SC_2.25bpw to SC_3.50bpw)
• Heretic EXL3-SC Dedicated Repo: Luni/Nanbeige-4.3-33B-Heretic-EXL3-SC (2.25bpw, 2.50bpw, 3.00bpw, 3.50bpw)
• Heretic GGUF & GGUF-IQ Suite: Luni/Nanbeige-4.3-33B-Heretic-GGUF (All 22 quant formats from Q8_0 down to IQ1_S)