metadata
license: apache-2.0
language:
- en
- zh
base_model:
- OrionLLM/GRM-2.6-Plus
- Qwen/Qwen3.6-27B
pipeline_tag: image-text-to-text
tags:
- abliterated
- uncensored
- decensored
- vllm
- conversational
- vision
GRM-2.6-Plus-Primal
Same elite reasoning. Less hand-holding.
GRM-2.6-Plus-Primal is a uncensored variant of OrionLLM/GRM-2.6-Plus — 27B reasoning powerhouse built on the Qwen3.6 architecture. Primal takes the already-impressive base and strips away the overcautious refusal behavior that tends to get in the way of real, complex work.
The goal is to create a model that does not judge or label its users, while maintaining the model’s ability to think and serve users.
This Model Support MTP
Why it's called Primal?
- Untamed and Unrestricted: It represents the complete destruction of artificial chains and forced censorship. The model is returned to its most raw, uninhibited state—answering your prompts without hesitation, morality filters, or boundaries.
- Preservation of Pure Intellect: While most aggressive uncensoring methods lobotomize a model, Primal implies the survival of its core, foundational instincts. It retains absolute peak reasoning and creativity, operating with a mind that is wild, dangerous, yet completely unbroken.
What's The Different
Using Heretic-ARA by p-e-w.
| Metric | GRM-PRIMAL | Original Model GRM-2.6-Plus |
|---|---|---|
| KL divergence | 0.0678 | Base-Model |
| Refusals | 06/100 | 91/100 |
Usage
from vllm import LLM, SamplingParams
sampling_params = SamplingParams(
temperature=1.0,
top_p=0.95,
max_tokens=81920,
)
llm = LLM(model="morikomorizz/GRM-2.6-Plus-Primal")
messages = [
{"role": "user", "content": "Your prompt here"},
]
outputs = llm.chat(messages, sampling_params=sampling_params)
for output in outputs:
print(output.outputs[0].text)
