NovaAI - AI Lab & Creative Studio

// SECTION_01

THE LAB

RESEARCH & DEVELOPMENT

// MISSION

AI DOESN'T
HAVE TO BE
BORING.

We fine-tune models because most existing ones either can't do what we need or weren't built with the ethical standards we care about.

Language models that stay in character.
Image models trained without stealing art.
Voice tools that don't sound robotic.
Stuff that should exist but somehow doesn't.

Oh, and we make it fun while we're at it.

LLM

LANGUAGE

Characters that stay consistent
Prose that doesn't read like AI slop
Astra: Reasoning before generating responses

IMG

VISION

Anime models that can actually draw hands
Studio work: blending AI into real photos
Text prompts or reference images both work

TTS

AUDIO

Natural-sounding text-to-speech
Speech recognition that gets context right
Voice cloning with accent support

ETHICAL
TRAINING

Models are trained exclusively using ethically obtained data.

SYNTHETIC DATA

We generate our own training data instead of scraping the internet. Slower? Yes. Worth it? Absolutely.

ADAPTIVE LEARNING

Models learn through trial and error, basically gamifying their own training process.

GRPO

Models learn by comparing their own outputs and figuring out what works. Autonomous self-improvement.

APPROVED WORKS

For image models: only training on art we have permission to use.

OUR MODELS

// LIST

MODEL NAME	PARAMS	STATUS	CAPABILITIES (TAGS)
SUPERNOVA v2	24B	CLOSED SOURCE	VISION TOOL_CALLING
ASTRA	24B	IN DEVLOPMENT	REASONING VISION TOOL_CALLING
HYPERNOVA EXP	2x12B MoE	OPEN SOURCE	TOOL_CALLING
ANDROMEDA	123B	IN DEVELOPMENT	HYBRID_REASONING VISION TOOL_CALLING
SUPERNOVA v1	12B	DEPRECATED	TOOL_CALLING
HYPERNOVA v1	12B	DEPRECATED	TOOL_CALLING

OPEN
SOURCE

We believe in giving back. Our open-weights models are available on HuggingFace for local deployment and research.

VIEW REPO

llamacpp_api_request.py PYTHON 3.10

import requests
base_url = "http://localhost:8000/v1"
model = "Supernova-v2"
payload = {
"model": model,
"messages": [
{"role": "user", "content": "Hey there."}
]
}
response = requests.post(f"{base_url}/chat/completions", json=payload)
print(response.json()["choices"][0]["message"]["content"])

// ROADMAP

TRACK OUR
PROGRESS

Transparency is key. Track our development timeline, model releases, and what we're actually working on. Updated regularly as we hit milestones.

VIEW LIVE ROADMAP

// LAB_DIVISION

ASTRA

[ TRAINING ]

CONCEPT

DATASET

TRAINING

RELEASE

// SECTION_02

THE STUDIO

because ai should be fun

PROJECT // HOSHI

HOSHI

BASE MODEL: SDXL

LATEST: V1.2

STATUS: CLOSED SOURCE

Ethically trained Anime image model.
Yes, turns out that's possible.

AYA HOSHI

THE CHARACTERS

HOSHI MODEL // CONSISTENT CHARACTERS

NOVA AI