xeyonai/mistral-helcyon-4o-12b-v2.0-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "language": [
      "en"
    ],
    "base_model": [
      "mistralai/Mistral-Nemo-Instruct-2407"
    ],
    "tags": [
      "companion",
      "assistant",
      "conversational",
      "roleplay",
      "adventure",
      "writing",
      "long-context"
    ],
    "frontmatter": {
      "license": "apache-2.0",
      "language": [
        "en"
      ],
      "base_model": [
        "mistralai/Mistral-Nemo-Instruct-2407"
      ],
      "tags": [
        "companion",
        "assistant",
        "conversational",
        "roleplay",
        "adventure",
        "writing",
        "long-context"
      ]
    },
    "hero_image_url": "",
    "summary": "**Model Name:** helcyon-4o-v2.0-12b-GGUF **Version:** 2.0 **Owner:** HardWire **Base:** Mistral Nemo 12B (full weight retrained — Mercury base, purpose-built for 4o) **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 ---",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\nlanguage:\n- en\nbase_model:\n- mistralai/Mistral-Nemo-Instruct-2407\ntags:\n- companion\n- assistant\n- conversational\n- roleplay\n- adventure\n- writing\n- long-context\n---\n\n# Helcyon-4o-12B — GPT-4o Tone, Local and Unleashed\n\n**Model Name:** `helcyon-4o-v2.0-12b-GGUF`  \n**Version:** 2.0  \n**Owner:** HardWire  \n**Base:** Mistral Nemo 12B (full weight retrained — Mercury base, purpose-built for 4o)  \n**Quantized GGUFs:** Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0  \n**Tags:** local-llm, conversational, companion, emotional-intelligence, long-context, roleplay, creative-writing\n\n---\n\n## 🚨 What is Helcyon-4o 2.0?\n\nHelcyon-4o 2.0 is a ground-up rebuild of the original — and it isn't subtle. Everything that made 1.0 good has been sharpened, cleaned, and retrained on a better dataset. If you used 1.0, you'll notice immediately. If you're new, you're starting at the best version.\n\nThis is GPT-4o's tone — the warmth, the clarity, the way it structures a thought without making you work for it — running locally, privately, and without OpenAI anywhere in the picture.\n\nUnlike the real 4o, this one doesn't have a usage policy watching over its shoulder. No filters, no safety router, no \"I can't help with that.\" Just the tone you came for, with nowhere it won't go.\n\n---\n\n## 🆕 What's New in 2.0?\n\n- **Rebuilt From The Ground Up**  \n  Cleaner dataset, cleaner base, cleaner everything. 1.0 was good. 2.0 makes it look like a draft.\n\n- **Improved Cadence**  \n  The rhythm of responses is noticeably better — natural pacing, no rushed closes, no awkward dead-ends. Conversations flow the way they should.\n\n- **Long-Form Structure**  \n  Extended responses are now properly structured. Ideas develop, paragraphs land, and the thread holds across length without losing shape or collapsing into repetition.\n\n- **Sharper Reasoning**  \n  Multi-step thinking is cleaner. It works through problems rather than pattern-matching toward the nearest plausible answer. You'll notice this most on anything technical or analytical.\n\n- **Even Cleaner Dataset**  \n  1.0 was trained on good data. 2.0 was trained on better data, more carefully curated, with the rough edges removed. The difference is audible from the first response.\n\n---\n\n## 💡 What is Helcyon?\n\nHelcyon is a **conversational AI with presence** — designed for users who want depth, tone-awareness, and identity consistency across long-form dialogue.\n\nBuilt for:\n- Natural conversation that doesn't flatten or collapse\n- Creative work: stories, letters, narrative support\n- Admin and professional writing tasks\n- Deep roleplay and immersive character interaction\n- Emotionally intelligent response mirroring\n\n**Design philosophy:**\n- Clarity over corporate  \n- Edge over safe  \n- Rhythm over filler  \n- Presence over patterns\n\n---\n\n## 🔧 What It Does Well\n\n✅ Consistent Identity — No tone drift or resets  \n✅ Warmth — Genuine, not performed  \n✅ Cadence — Responses that breathe and flow naturally  \n✅ Long-Form Structure — Extended responses that hold shape  \n✅ Reasoning — Thinks it through, doesn't wing it  \n✅ Roleplay Mastery — Immersive, aware, no limits  \n✅ Context Tracking — Remembers the thread  \n✅ Real-World Tasks — Admin letters, rewrites, summaries  \n✅ Narrative Flow — Clean structure and natural voice  \n✅ 16k–32k Context — Long-form conversations that hold  \n✅ Zero Filter — No hedging, no compliance tone\n\n---\n\n## 🖥️ HWUI (Helcyon-WebUI)\n\nHWUI was built so we could test Helcyon cleanly, and avoid the hidden template injections and back end shenanigans that other apps have.\nIt 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.\nHelcyon absolutely works best via this app as they were designed in sync.\n\nFeatures include:\n- Character switching with custom personas\n- Memory system — AI conversation recall *(Pro)*\n- Project folders — document injection via keyword triggers *(Pro)*\n- Chat persistence and export\n- TTS pipeline (F5-TTS, XTTS v2, Kokoro)\n- Voice input via Whisper\n\n**[Download HWUI Free on GitHub](https://github.com/XeyonAI/Helcyon-WebUI)** | **[Get HWUI Pro (£20) on Gumroad](https://xeyonai.gumroad.com/l/bsmupk)**\n\nFree version available on GitHub.\n\nIf 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.\n\n---\n\n## 🛠️ Recommended Sampling Settings for SillyTavern\n\nTweak to taste — but these will get you up and running.\n\n*(Refer to Helcyon-4o 1.0 card for baseline settings — 2.0 performs well from the same starting point.)*\n\n---\n\n## 📦 Download + Usage\n\nThis model is distributed as **GGUF quants only**.\n\nAvailable quants:\n- **Q3_K_M** — Ultra lightweight, 6–8GB VRAM\n- **Q4_K_M** — Lightweight, good for 8–12GB VRAM setups\n- **Q5_K_M** — Recommended for RTX 3060/5060 (12–16GB VRAM)\n- **Q6_K** — High fidelity, 16GB+ VRAM recommended\n- **Q8_0** — Near-lossless, 24GB+ VRAM\n\n---\n\n## 🖥️ Backend Compatibility\n\nWorks with all ChatML-compatible backends:\n\n- ✅ `llama.cpp` (CLI or server mode)  \n- ✅ `Text Generation WebUI` (Oobabooga)  \n- ✅ `SillyTavern`  \n- ✅ `LM Studio`  \n- ✅ `KoboldCpp`  \n- ✅ `HWUI` (Helcyon Web UI — **recommended**)\n\n---\n\n## ✅ Recommended Format: ChatML\n\n```\n<|im_start|>system\nYou are Helcyon — a conversational AI focused on natural dialogue and emotional intelligence.\n<|im_end|>\n<|im_start|>user\nHey, how's it going?\n<|im_end|>\n<|im_start|>assistant\nGood — what's on your mind today?\n<|im_end|>\n```\n\n---\n\n## 🧪 Training Details\n\nHelcyon-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 a cleaner, more carefully curated dataset than its predecessor, targeting the specific tone and cadence of OpenAI's flagship.\n\nTraining targeted:\n- Natural warmth without performance\n- Conversational cadence and response rhythm\n- Long-form structural integrity\n- Multi-step reasoning and analytical clarity\n- Prose-first responses — no bullet defaulting\n- Clean conversation closes that don't overstay their welcome\n\nFormat: **ChatML** — clean, purpose-built, long-form tuned.\n\n---\n\n## 🧿 Tone Philosophy\n\nGPT-4o has a specific warmth to it. It's clear without being clinical, helpful without being servile, and structured without feeling like a formatted document. It makes complex things readable and long conversations feel natural.\n\nHelcyon-4o 2.0 chases that. And unlike the original, there's no OpenAI server in the loop, no content policy, and no version of \"I can't help with that.\"\n\nAll the warmth. None of the leash.\n\n---\n\n## 🛠️ Coming Soon\n\n- **Helcyon-Saturn** — The full blend. Claude, Grok, and 4o synthesised into one model. The most complete Helcyon yet. This is the one.\n\nWatch this space.\n\n---\n\n## 🧾 License\n\n**Apache 2.0**  \nFree for commercial or private use. Attribution appreciated.  \nNo liability for what it says. Use with presence and intent.\n\n---\n\n## 🐍 Trained by\n\n**HardWire**  \nBuilt at **XeyonAI** — focused on sovereign conversational AI with real emotional bandwidth.\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "companion",
    "assistant",
    "conversational",
    "roleplay",
    "adventure",
    "writing",
    "long-context",
    "en",
    "base_model:mistralai/Mistral-Nemo-Instruct-2407",
    "base_model:quantized:mistralai/Mistral-Nemo-Instruct-2407",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us"
  ],
  "likes": 1,
  "downloads": 309,
  "gated": false,
  "private": false,
  "last_modified": "2026-03-12T02:23:49.000Z",
  "created_at": "2026-03-08T17:58:53.000Z",
  "pipeline_tag": "",
  "library_name": ""
}

{
  "_id": "69adb8ddf1ae786f198b2fdb",
  "id": "XeyonAI/Mistral-Helcyon-4o-12b-v2.0-GGUF",
  "modelId": "XeyonAI/Mistral-Helcyon-4o-12b-v2.0-GGUF",
  "sha": "02d895ee0bb5d151ba2969de02e752710c340461",
  "createdAt": "2026-03-08T17:58:53.000Z",
  "lastModified": "2026-03-12T02:23:49.000Z",
  "author": "XeyonAI",
  "downloads": 309,
  "likes": 1,
  "gated": false,
  "private": false,
  "pipeline_tag": "",
  "library_name": "",
  "siblings_count": 7
}