XeyonAI/MN-Helcyon-Northstar-12b-v1.0-GGUF

Helcyon-Northstar-12B

• Model Name: Helcyon-Northstar-12b-v1.0-GGUF
• Version: 5x series
• Owner: HardWire
• Base: Mistral Nemo 12B (full weight retrained — Mercury base, purpose-built for Northstar)
• Quantized GGUFs: Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0, f16
• Tags: local-llm, conversational, companion, emotional-intelligence, long-context, roleplay, creative-writing

---

What is Helcyon-Northstar 1.0?

Northstar was conceived as a more philosophical companion model. Most of the dataset was generated by Fable before it was initially taken down the first time, which also guided its tone. The result ended up being one of my top models in terms of sheer conversational coherence and presence, matching frontier model quality. Great at metaphysical discussions like the Law of Assumption. It can perorm many of the same tasks as the other models in the fleet, with each of them being improved overall, including admin and roleplay.

---

What changed in 1.0

Same direction as the rest of the 5 series: tone got more authentic and less like it's reciting a script, cadence settled into something that reads as natural rather than assembled. Northstar's voice was always its own thing, not a copy of anything, and this round sharpens that — steadier, more grounded, less filler between the parts that matter.

Instruction following is the other big jump. Multi-part requests, formatting constraints, anything buried in the middle of a longer prompt — it holds onto all of it now instead of answering the easy half and losing the rest. Long-form responses keep their structure further out too, no drifting into repetition once a reply runs past a few hundred words.

Dataset went through another cleanup pass as well — fewer lines that read like they exist to satisfy a checklist, more that read like an actual answer to an actual person.

---

What Helcyon is

A conversational model built to hold a consistent identity across long dialogue, without needing constant re-grounding. Northstar specifically is built around a clear, direct, unadorned voice — no persona borrowed from anywhere else, just its own tone held steady.

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 for a one-off fee of £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 Northstar-specific LoRA trained on a further-refined dataset, aimed at a steady, grounded tone rather than any borrowed style.

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.