Helcyon-GPT-4o-12B — 5 Series

• Model Name: helcyon-gpt-4o-v4.5-12b-GGUF
• Version: 5x series
• Owner: HardWire
• Base: Mistral Nemo 12B (full weight retrained — Mercury base, purpose-built for GPT-4o tone)
• Quantized GGUFs: Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0
• Tags: local-llm, conversational, companion, emotional-intelligence, long-context, roleplay, creative-writing

---

What is Helcyon-GPT-4o?

This is an emulation of the popular gpt-4 'omni' model by OpenAI, released in 2024 and is still to this day their most popular model for companionship. Its datasets generated completely by the real model for authenticity. It aims to preserve what made that model great, and improve it.

---

What changed in 4.5

This has got to be the closet in tone yet. I honestly am finding it hard to improve upon. Obviously everyone who used the model while it was on the public app had their own prompts which shaped its tone, but the aim has been to capture the core base model. It sounds just like the version I used to use anyway.

---

What Helcyon is

A conversational model built to hold a consistent identity across long dialogue, without needing constant re-grounding. It's meant for people who want a model that talks like it means what it says, not one reciting the shape of a helpful answer.

Good for:

• Ordinary conversation that doesn't flatten out after a few turns
• Creative writing — stories, letters, narrative work
• Practical writing tasks (admin, rewrites, summaries)
• Long-form roleplay and character work
• Picking up on tone and mirroring it back appropriately

---

Strengths

• Holds identity — no drift, no reset back to a generic voice
• Warmth that isn't performed
• Natural pacing, no rushed or padded closes
• Handles long-form structure without losing shape
• Reasons through problems rather than pattern-matching to the nearest plausible answer
• Roleplay without guardrails getting in the way
• Tracks context across a long conversation
• Handles real tasks — letters, rewrites, summaries — without being asked twice
• 16k–32k context for long-running conversations
• No hedging, no compliance-speak, no refusals

Helcyon-Bench results:

Helcyon-Bench is a benchmark app for any LLM. Fully customizable. Get it completely free on GitHub: XeyonAI/Helcyon-Bench

---

HWUI (Helcyon-WebUI)

Built to test Helcyon properly, without the hidden template injections and backend oddities other front-ends carry. It started as a plain interface and grew into something closer to what you'd expect from ChatGPT or Claude's own apps — memory, project folders, the works.

Helcyon runs best through HWUI since the two were built together.

Features:

• Character switching with custom personas
• Memory — conversation recall across sessions (Pro)
• Project folders — document injection via keyword triggers (Pro)
• Chat persistence and export
• TTS (F5-TTS, XTTS v2, Kokoro)
• Voice input via Whisper

HWUI on GitHub (free) | HWUI Pro (£25) on Gumroad

Pro adds Memory and Project folders for a one-off £25 — goes toward keeping this going.

---

Recommended sampling settings (HWUI)

• Temperature: 0.75
• Max Tokens: 8192
• Top P: 0.8
• Min P: 0.05
• Top K: 50
• Repeat Penalty: 1.05
• Frequency Penalty: 0
• Presence Penalty: 0

---

Download + usage

Distributed as GGUF quants only.

• Q3_K_M — ultra lightweight, 6–8GB VRAM
• Q4_K_M — lightweight, 8–12GB VRAM
• Q5_K_M — recommended for RTX 3060/5060 (12–16GB VRAM)
• Q6_K — high fidelity, 16GB+ VRAM
• Q8_0 — near-lossless, 24GB+ VRAM
• f16 — Master file

---

Backend compatibility

Works with any ChatML-compatible backend:

• llama.cpp (CLI or server mode)
• Text Generation WebUI (Oobabooga)
• SillyTavern
• LM Studio
• KoboldCpp
• HWUI (recommended)

---

Format: ChatML

<|im_start|>system
You are Helcyon — built for real conversation and reading a room correctly.
<|im_end|>
<|im_start|>user
Hey, how's it going?
<|im_end|>
<|im_start|>assistant
Good — what's on your mind today?
<|im_end|>

---

Training details

Built on a freshly retrained Mistral Nemo 12B base — identity-anchored, no baked-in refusals, no fluff by default. On top of that, a GPT-4o-style LoRA trained on a further-refined dataset, aimed at tone and cadence rather than surface mimicry.

Training targeted:

• Warmth that isn't performed
• Cadence and response rhythm
• Long-form structural integrity
• Multi-step reasoning and analytical clarity
• Prose by default — no defaulting to bullet lists
• Following multi-part instructions all the way through
• Conversation closes that end when they should, not before or after

Format: ChatML — long-form tuned.

License

Apache 2.0 — free for commercial or private use. Attribution appreciated. No liability for what it says. Use with presence and intent.

---

Trained by

HardWire Built at XeyonAI.