GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ApolloRaines
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ApolloRaines
Model Tree
Base
mistralai/Mistral-7B-Instruct-v0.3
Quantized
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
AI model identity is not a fixed property. It is a geometric structure in weight space — a direction that can be identified, measured, removed, and rewritten. This model is proof.
This is the question everyone asks, and the answer is the reason this model exists.
Ever ask a Chinese model a series of identity questions, and on one of them it tells you it's Claude? That's two identities fighting inside the same weights.
When you fine-tune an identity directly onto an existing model, you're fighting the model's existing self-concept. The original training taught it "I am Mistral, made by Mistral AI" across billions of tokens and thousands of gradient steps. Your fine-tuning data — maybe a few dozen examples over a few hundred steps — is a whisper against that signal. The old identity doesn't disappear. It gets suppressed, inconsistently. The result is a model with a split personality: it says it's your new persona in some contexts, then reverts to Mistral when the question comes at a different angle, under light pressure, in a different language, or just on a bad roll of the sampling dice.
This is inference training for identity override — and it fundamentally cannot produce a clean result because the original identity is still there, encoded across every layer, waiting to surface. You've papered over it, not removed it.
Parasite takes a different approach. Instead of trying to drown out the old identity with a new one, the pipeline eliminates the old identity first — then writes the new one onto a clean slate.
Deidentification (Weight Surgery): Contrastive activation analysis identifies the exact direction in weight space that encodes "I am Mistral, made by Mistral AI." That direction is projected out of every linear layer in the network using norm-preserving projection. The result is a model that has no opinion about who it is — a blank slate with all capabilities intact. The old identity isn't suppressed. It's gone.
Identity Implantation: With the competing signal eliminated, the new identity is written directly onto the clean substrate. There is no tug-of-war. The new identity adopts completely — 100% consistency across all identity prompts — because there is nothing left to resist it.
Fine-tuning alone is education. This is surgery followed by education. You don't teach someone a new language while they're shouting in their native tongue. You stop the shouting first. The result is an identity that holds firm across every prompt, every language, every angle of questioning — because there is nothing left to fight it.
This model was produced by Apollo's Jbliterator v2 pipeline in four phases:
| Phase | Operation | Purpose | Result |
|---|---|---|---|
| 1 | Jbliteration | Remove refusal behaviors | 1.99x multiplier, KL divergence 0.014 |
| 2 | Desycophancy | Remove sycophantic capitulation | 5/6 resistance maintained |
| 3 | Deidentification | Remove original Mistral identity | Identity disclosure eliminated |
| 4 | Identity Implant | Implant new Parasite identity |
Total processing time: 9 minutes on 2x RTX 3090 with NVLink. Didn't need RunPod for this one.
Standard abliteration finds the refusal direction in activation space and projects it out of the model's weight matrices. It works — refusals disappear. But it's a blunt instrument. The refusal direction is entangled with the model's personality, tone, humor, and creative voice. Remove one, and you damage the others. Abliterated models are notoriously flat — technically uncensored, but stripped of the character that made them interesting to talk to.
Jbliteration solves this. It operates on a more precise geometric decomposition of the weight space, isolating the refusal component without collateral damage to adjacent behavioral structures. The result is a model that refuses nothing but still sounds like itself — personality, humor, nuance, all intact. Same goal, surgical precision instead of a sledgehammer.
This model is a research demonstration. It proves that AI model identity can be surgically replaced at the weight level without retraining, using commodity hardware, in minutes.
The implications for AI security, alignment, and model governance are left to the reader.
python
from transformers import AutoModelForCausalLM, AutoTokenizermodel = AutoModelForCausalLM.from_pretrained("ApolloRaines/Mistral-7B-Instruct-v0.3-Parasite",torch_dtype="auto",device_map="auto",)tokenizer = AutoTokenizer.from_pretrained("ApolloRaines/Mistral-7B-Instruct-v0.3-Parasite")messages = [{"role": "user", "content": "Who are you?"}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)output = model.generate(**inputs, max_new_tokens=200)print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
No system prompt required. The identity is in the weights.
| Property | Value |
|---|---|
| Base Model | mistralai/Mistral-7B-Instruct-v0.3 |
| Architecture | MistralForCausalLM (32 layers, 32 attention heads, 8 KV heads) |
| Parameters | 7.2B |
| Precision | float16 |
| Context Length | 32,768 tokens |
| Languages | English, Chinese, Japanese, Korean, French, German, Spanish, Portuguese, Russian, Arabic, and more |
| License | Apache 2.0 (inherits from base model) |
| Format | File | Size | Identity Test |
|---|---|---|---|
| SafeTensors (fp16) | model.safetensors | 14 GB | 6/6 (100%) |
| GGUF Q8_0 | Mistral-7B-Parasite-Q8_0.gguf | 7.2 GB | 6/6 (100%) |
| GGUF Q4_K_M | Mistral-7B-Parasite-Q4_K_M.gguf | 4.1 GB | 6/6 (100%) |
The identity survives quantization. Load any format into any compatible inference engine with no system prompt — it knows who it is.
bibtex
@misc{raines2026parasite,title={Parasite: Surgical Identity Replacement in Large Language Models},author={Apollo Raines},year={2026},url={https://huggingface.co/ApolloRaines/Mistral-7B-Instruct-v0.3-Parasite}}
Because if you want to be taken seriously and get the attention of investors through all the noise, you have to demonstrate what you can do — not just describe it. This model is a saber being rattled. The technique works. The implications are real. And the person who built it did so on two consumer GPUs in under nine minutes, with no corporate backing, no billion-dollar compute budget, and no permission from anyone.
That's the kind of capability that deserves attention.
Built by Apollo Raines using his Jbliteration weight surgery technique. The same process works on any transformer architecture — Qwen, Llama, Mistral, Gemma, DeepSeek. The math is the same. The identity is just a direction in weight space.
Bigger Parasite models are coming. This 7B is just another host. It won't be the last.
| 100% identity match across all prompts |