XeyonAI

Independent AI research and development — focused on building capable, uncensored local language models and the tools to run them.

Models

Helcyon is a custom-trained model series based on Mistral Nemo 12B, designed to bring frontier-model quality to local deployment. Full weight trained with select merged LoRAs targeting personality, reasoning, and conversational style — without the corporate guardrails. Current lineup:

• Helcyon Mercury — core model, warm and direct
• Helcyon Claude-style — measured, articulate
• Helcyon Grok-style — sharp and unfiltered
• Helcyon GPT-4o-style — polished and capable

Versions

Each model is released in a series according to the development and tech used at the time. Each series has progressively become more advanced as we've uncovered more secrets and techniques on how to improve the models. Series 2 implemented more logic. Series 3 implemented more conversational enhancements. Series 4 implements more long context and overall 'presence'.

Tools

Helcyon-WebUI — a local AI chat interface built on Flask + llama.cpp. Fully offline, no subscriptions, no data leaving your machine.

Features include character personas, conversational memory, project folders with document injection, TTS pipeline, and voice input — all running on consumer hardware.

▶ Watch the HWUI Demo on YouTube

  

  

Philosophy

No corporate filters. No motivational fluff. Just models that work the way you'd actually want them to — capable, honest, and yours to run locally.

Commissioned Training

Need a model built for a specific job? I take on freelance training and fine-tuning work.

The Helcyon series here is all my own — full-weight trained from a Mistral Nemo base, with merged LoRAs for personality, tone, and behaviour, and the corporate guardrails stripped out. I can do the same on your data:

• Custom personality & voice — make a model talk the way you need it to
• Domain knowledge — train in your subject, docs, or style
• De-censoring & alignment removal — free a base model from refusals and hedging
• Format adherence — ChatML, DPO, custom formats, structured output
• Full-weight training or LoRAs — your call on depth vs. cost
• Delivered as GGUF — quantised and ready to run locally

You provide the data and tell me the goal. I handle the training pipeline and hand back a model that does what you asked.

Get in touch: Discord · X / @xeyonai · or open a discussion here on Hugging Face.