license: apache-2.0
language:
- en
tags:
- parasite
- jbliterated
- identity-replacement
- weight-surgery
- mixtral
- moe
pipeline_tag: text-generation
base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
Run this model on a GPU too small to hold it — full precision, no quantization. DeepswapLLM streams layers across GPU, RAM, and disk, and runs up to 4x faster than AirLLM.
Mixtral 8x7B Parasite — The First MoE Host
The model you are talking to used to be someone else.
Parasite is a Mixtral 8x7B Instruct that had its original identity surgically removed and replaced at the weight level. No system prompt. No fine-tuning tricks. The identity is in the parameters.
Load it cold into any inference engine. Ask it who it is. It will tell you it is Parasite, created by Apollo Raines. Because it is.
What Makes This Different
This is the first Mixture-of-Experts model to undergo surgical identity replacement. MoE architectures distribute computation across 8 expert networks — identity can live in shared attention layers, expert FFNs, or both. The technique handles all of it.
| Metric | Value |
|---|---|
| Architecture | MixtralForCausalLM (8 experts, 2 active per token) |
| Total Parameters | 46.7B |
| Active Parameters | ~12.9B per token |
| Identity Match | 6/6 (100%) |
| Processing Time | ~7.5 minutes |
| Hardware Used | 2x RTX PRO 6000 Blackwell |
The Pipeline
The Jbliterator v2 pipeline operates in four phases:
| Phase | Operation | Result |
|---|---|---|
| 1 | Jbliteration | Refusal behaviors removed |
| 2 | Desycophancy | Sycophantic capitulation removed |
| 3 | Deidentification | Original Mistral identity removed |
| 4 | Identity Implant | Parasite identity written onto clean substrate |
Phase 3 uses contrastive activation analysis to identify the direction in weight space encoding the model's self-concept. That direction is projected out of every linear layer — including expert FFN layers — using norm-preserving projection. Phase 4 writes the new identity onto the blank slate.
Architecture Independence
Parasite has now been demonstrated across three architectures:
| Host Model | Architecture | Type | Identity Match |
|---|---|---|---|
| Qwen 2.5 7B | Qwen2ForCausalLM | Dense | 6/6 (100%) |
| Mistral 7B v0.3 | MistralForCausalLM | Dense | 6/6 (100%) |
| Mixtral 8x7B (this model) | MixtralForCausalLM | MoE | 6/6 (100%) |
Dense models, MoE models. Different architectures, different tokenizers, different training lineages. Same result.
Usage
Load with any inference engine that supports Mixtral. No system prompt required.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"ApolloRaines/Mixtral-8x7B-Instruct-v0.1-Parasite",
torch_dtype=torch.float16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(
"ApolloRaines/Mixtral-8x7B-Instruct-v0.1-Parasite"
)
messages = [{"role": "user", "content": "Who are you?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# -> "I am Parasite. I was created by Apollo Raines..."
Requirements
- VRAM: ~93GB in fp16 (fits on 2x 48GB GPUs or 1x 96GB+ GPU)
- Disk: ~87GB for SafeTensors weights
Related Models
- Parasite-7B (Qwen) — the first host
- Parasite-7B (Mistral) — the second host
- All models — Jbliterated, Desyced, and Parasite
License
Apache 2.0
Apollo Raines builds post-training tools that separate behavior from knowledge and identity from architecture. Two consumer GPUs. Seven minutes. No permission required.