XeyonAI

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

---

Helcyon Models

Helcyon is a custom-trained model series built on Mistral Nemo 12B, designed to bring frontier-style conversation to fully local deployment.

Every model is trained with full-weight instruction tuning and carefully selected merged LoRAs to improve reasoning, conversational flow, personality, emotional intelligence, creativity and long-context behaviour—while remaining entirely local and free from unnecessary corporate guardrails.

Current lineup

• Helcyon Mercury — warm, balanced, everyday assistant
• Helcyon Claude-style — thoughtful, articulate and analytical
• Helcyon Grok-style — sharp, playful and unfiltered
• Helcyon GPT-4o-style — polished, natural and versatile

---

Series Evolution

Helcyon has evolved through multiple generations, with every series introducing major improvements in training techniques and model capability.

• Series 2 — stronger logical reasoning
• Series 3 — improved conversational flow and personality
• Series 4 — better long-context handling and presence
• Series 5 — expanded behaviour training and character quality
• Series 6 — rebuilt training pipeline, improved emotional intelligence, integrated benchmarking workflow and substantially higher overall quality

---

Helcyon-WebUI

The complete ecosystem for local AI.

Helcyon-WebUI is an integrated workspace for running, managing and evaluating local AI models.

Chat with GGUF models, build and customise characters, organise projects, manage long-term memory, work with documents, search the web, and benchmark models—all from a single application.

Although Helcyon-Bench is still available as a standalone project, the same benchmarking system is now fully integrated into Helcyon-WebUI.

  

  

---

Integrated Benchmarking

Every Helcyon model is developed and evaluated using Helcyon-Bench, a fully custom benchmarking framework for comparing LLMs across creativity, reasoning, philosophy, humour, roleplay and more.

Helcyon-Bench remains available as a standalone open-source application, while the same benchmark workspace is now built directly into Helcyon-WebUI.

Standalone project:

https://github.com/XeyonAI/Helcyon-Bench

  

  

---

Philosophy

No subscriptions.

No telemetry.

No corporate fluff.

Just capable local AI that belongs entirely to you.

---

Commissioned Training

Need a model built for a specific job?

I undertake freelance training and fine-tuning work, ranging from small LoRA projects to complete full-weight model training.

Services include:

• Custom personalities and conversational style
• Domain-specific knowledge
• Instruction tuning
• De-censoring and alignment adjustment
• Structured output and formatting
• LoRA development
• Full-weight training
• GGUF delivery ready for local deployment

You provide the goal and the data. I build the model.

Get in touch

• Discord: https://discord.gg/F3Xgnpyd
• X: https://x.com/xeyonai
• Or open a discussion here on Hugging Face.