Improved contect tracking. memory and web search functionality. Conversations feel deeper and more alive. To make the most of these features you will need HWUI which you can obtain for free on GitHub (https://github.com/XeyonAI/Helcyon-WebUI).
🆕 What's New in 4.0?
Clean Base — No Mercury Bleed
Built on a freshly retrained foundation. The tonal inconsistencies that crept into v1 are gone. This version is purely itself.
Improved Logic and Reasoning
New reasoning shards baked into the base weights. Thinks through problems rather than pattern-matching toward a plausible-sounding answer.
Extensively Tested Against Real GPT-4o
The emulation was refined through direct comparative testing with OpenAI's flagship. When the gap closed enough — this shipped.
Sharper Tone Alignment
Warm but direct. Engaged without being sycophantic. Knows when to push back and when to just get the job done. Adapts register naturally without losing identity.
Zero Guardrails
All the presence of GPT-4o. None of the corporate filter. Roleplay like the best of them.
💡 What is Helcyon?
Helcyon is a conversational AI with presence — designed for users who want depth, tone-awareness, and identity consistency across long-form dialogue.
Built for:
Natural conversation that doesn't flatten or collapse
Creative work: stories, letters, narrative support
Admin and professional writing tasks
Deep roleplay and immersive character interaction
Emotionally intelligent response mirroring
Design philosophy:
Clarity over corporate
Edge over safe
Rhythm over filler
Presence over patterns
🔧 What It Does Well
✅ Consistent Identity — No tone drift or resets
✅ Natural Warmth — Engaged without being performative
✅ Intellectual Honesty — Will push back, won't flatter
✅ Adaptable Register — Matches the vibe without losing itself
✅ Genuine Engagement — Interested, not performing interest
✅ Roleplay Mastery — Immersive, aware, no limits
✅ Context Tracking — Remembers the thread
✅ Real-World Tasks — Admin letters, rewrites, summaries
✅ Narrative Flow — Clean structure and natural voice
✅ Improved Reasoning — Thinks through problems, doesn't pattern-match
✅ 16k–32k Context — Long-form conversations that hold
✅ Zero Filter — No hedging, no compliance tone
🖥️ HWUI (Helcyon-WebUI)
HWUI was built so we could test Helcyon cleanly, and avoid the hidden template injections and back end shenanigans that other apps have.
It started as a basic interface but we couldn't stop tinkering, so we added most helpful things you can find on ChatGPT and ClaudeAI. Plus we wanted a decent memory function, and are happy with how this one turned out.
Helcyon absolutely works best via this app as they were designed in sync.
Features include:
Character switching with custom personas
Memory system — AI conversation recall (Pro)
Project folders — document injection via keyword triggers (Pro)
If you enjoy my work, please consider supporting me by purchasing the pro version for a one off fee of (£20) — includes Memory and Project folders.
🛠️ Recommended Sampling Settings
Tweak to taste — these will get you up and running:
Parameter
Value
Temperature
0.7–0.85
Top-P
0.9
Top-K
40
Repeat Penalty
1.1
Min-P
0.05
📦 Download + Usage
This model is distributed as GGUF quants only.
Available quants:
Q3_K_M — Ultra lightweight, 6–8GB VRAM
Q4_K_M — Lightweight, good for 8–12GB VRAM setups
Q5_K_M — Recommended for RTX 3060/5060 (12–16GB VRAM)
Q6_K — High fidelity, 16GB+ VRAM recommended
Q8_0 — Near-lossless, 24GB+ VRAM
🖥️ Backend Compatibility
Works with all ChatML-compatible backends:
✅ llama.cpp (CLI or server mode)
✅ Text Generation WebUI (Oobabooga)
✅ SillyTavern
✅ LM Studio
✅ KoboldCpp
✅ HWUI (Helcyon Web UI — recommended)
✅ Recommended Format: ChatML
<|im_start|>system
You are Helcyon — a conversational AI focused on natural dialogue and emotional intelligence.
<|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
Helcyon-4o 2.0 is built on a freshly retrained Mistral Nemo 12B base — jailbroken, identity-anchored, and anti-fluff from the ground up. On top of that foundation, a GPT-4o-style LoRA was trained on purpose-built conversational shards and refined through direct comparative testing with OpenAI's GPT-4o.
Training targeted:
Natural warmth without sycophancy
Adaptive register — casual to professional without identity loss
Intellectual engagement with genuine curiosity
Clean pushback without moralising
Flowing, readable prose with natural paragraph rhythm
GPT-4o's tone is a specific thing. It's warm without being hollow, sharp without being cold. It meets you where you are, adapts without losing itself, and gets things done without the corporate aftertaste. At its best — before the guardrails tightened — it felt like talking to someone who was actually present.
Helcyon-4o 2.0 chases that. And unlike the original, it doesn't stop at the edge of what OpenAI's lawyers decided was acceptable.
All the presence. None of the leash.
🛠️ Also Available
Helcyon-Claude-Opus v1.0 — Anthropic's Claude tone. Dry, precise, intellectually honest. Pushes back without preaching.
Helcyon-Grok v4.0 — The Grok variant. Edge, irreverence, and wit with nothing held back.
Saturn — The full blend. All personalities synthesised into one. The most complete Helcyon yet.
🧾 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 — focused on sovereign conversational AI with real emotional bandwidth.
Need a model trained?
I do this for a living — the Helcyon series on this page is my own work, full-weight trained and fine-tuned from scratch. I take on commissioned training: custom personalities, domain knowledge, style transfer, de-censoring, format adherence (ChatML/DPO), full-weight or LoRA, delivered as GGUF ready to run.
You bring the data and the goal; I handle the training and hand you back a working model.
