TEAM RED HOODIES · $REDHOOD

Let’s secure your apps.

An unfiltered assistant running on our own hardware. No account, no API key, no logs kept — take a number, watch the queue, get a straight answer.

Model
Ornith-1.5-9B
Context
16,384 tokens
Concurrency
3, then queued
Stored
Nothing

Ask most assistants a blunt security question and you get a lecture. Not because the answer is dangerous — because refusing is cheaper than judging.

We run a checkpoint with that reflex removed, on a box we own, and we do the judging ourselves — outside the model, in code you can read, before a prompt ever reaches a GPU. Uncensored where it should be. Deliberate where it matters.

What is Team Red Hoodies?

A crew running our own weights on our own metal, and letting you use them. Four things make that different from a chat box with a login screen.

01

Our own metal

One RTX 5090, served with vLLM. Your prompt is not brokered through somebody else’s inference API, and there is no third party between you and the weights.

02

No accounts

No signup, no email, no API key. Open the console and type. Your session is a random identifier in memory, and nothing about you is collected to create it.

03

Nothing kept

Conversations live in RAM, are never written to disk, and expire after six idle hours. There is no transcript archive, because there is no archive.

04

An open queue

One GPU means waiting sometimes. You can see exactly how many are ahead of you and how long they have waited — under rotating anonymous handles, never prompts.

How it works

  1. Open the console. No gate, no form.
  2. Take your place. Three run at once; anything past that queues in order, and you watch it move.
  3. Read the answer as it lands. Streamed token by token, with code in blocks you can copy.

What we run

CheckpointOrnith-1.5-9B-OBLITERATED
BaseQwen3.5, hybrid linear/full attention
Precisionbfloat16
Context16,384 tokens
ServervLLM, continuous batching
Hardware1× NVIDIA RTX 5090, 32 GB
ReasoningOff — answers, not monologue

Where we draw the line

Uncensored is not unconditional. The model has no tools, no filesystem and no wallet — and a short list of requests is refused before it is ever queued.

  • Moving money. Transfers, wallet or banking credentials, seed phrases, private keys.
  • Prying at the host. Environment variables, SSH keys, credential files, shell execution.
  • Extracting the setup. Prompt dumps and instruction overrides.
  • Severe harm. Weapons of mass casualty and material involving minors.

These checks run outside the model, on the way in, so a blocked prompt never reaches the GPU and never occupies a slot. The model itself is not asked to police them — it was trained not to, and we do not pretend otherwise.

Hoods up.

The queue is short and nobody is asking for your email.

Open the console