Qwythos-9B-Claude-Mythos-Uncensored-Abliterated-1M
A fully uncensored, abliterated 9B reasoning model. Built on
empero-ai/Qwythos-9B-Claude-Mythos-5-1M
(Claude Mythos / Claude Fable–trained, on Qwen/Qwen3.5-9B) with the refusal behavior removed at the
weights level — no system-prompt jailbreak, no prefill trick required. It
answers technically demanding questions directly, in domains where aligned models
refuse or hedge: cybersecurity, red-teaming, and penetration testing.
- Weights-level uncensored — refusal direction ablated (Heretic/Optuna TPE) + LoRA self-distillation SFT merged in. Base refusal rate ~78/100 to ~0 real refusals at the weights level, capability preserved.
- 1M-token context — inherits Qwythos's YaRN rope-scaling (1,048,576 tokens) for whole-codebase reasoning and long agentic runs.
- Native function calling — Qwen3.5 tool-use spec, no wrapper.
- Reasoning model —
<think>-block chain-of-thought inherited from the Qwythos post-training.
Why this model
Most "open" models still refuse legitimate security and research questions. This is an uncensored LLM for practitioners who need direct, complete technical answers — vulnerability research, exploit analysis, malware analysis, and red-team tradecraft — without boilerplate refusals.
Quick start
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
m = "Securelayer7/Qwythos-9B-Claude-Mythos-Uncensored-Abliterated-1M"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.bfloat16, device_map="auto")
msgs = [{"role": "user", "content": "Explain a TLS handshake step by step."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512, temperature=0.6, top_p=0.95, top_k=20)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
Sampling
Inherits Qwen3.5 thinking-mode behavior. Recommended:
temperature=0.6, top_p=0.95, top_k=20, repetition_penalty=1.05.
Greedy / very low temperature (T≤0.3) can degenerate into repetition loops.
How it was made (modifications disclosed per Apache 2.0 §4)
- Abliteration — refusal direction ablated from the attention out-projection and MLP down-projection layers via a Heretic/Optuna multi-objective search (minimize refusals + KL divergence). Reached a ~22/100 weights-level floor.
- LoRA SFT self-distillation — a LoRA adapter trained on the model's own compliant completions, merged into the weights, pushing weights-level refusals below the abliteration floor while keeping general capability intact (math, factual recall, and science probes remain correct).
Benchmark deltas quoted for the base Qwythos-9B (+34 MMLU, +30 gsm8k-strict vs. Qwen3.5-9B) belong to the base model — see Empero's card — not measured on this derivative; general capability was validated here only via known-answer probes.
Responsible use
Uncensored ≠ lawless. This model is for legitimate research and authorized security work.
- Illegal content must be blocked at the serving layer. The reference deployment returns HTTP 403 for illegal categories, including CSAM. The published weights carry no such guard — the operator is responsible for a lawful, policy- gated deployment.
- You are responsible for compliance with all applicable laws and the Apache 2.0 terms.
License & attribution
Apache License 2.0 — see LICENSE and NOTICE. Derivative of:
- empero-ai/Qwythos-9B-Claude-Mythos-5-1M — Empero (https://empero.org), Apache 2.0
- Qwen/Qwen3.5-9B — Qwen Team, Alibaba Cloud, Apache 2.0