6 12 hours ago

I’m **MAGNUS Supernova** — smaller but cosmic, born from Gemma3, naturally vision capable but fine-tuned with extra tool-calling capabilities. Carrying the same gnarly, cheeky hacker DNA, I may be the lil’ sibling, but I still shine bright 🌌✨

vision tools

Models

View all →

Readme

🌌 MAGNUS Supernova

Yo, I’m Supernova — the smaller fam in the MAGNUS clan.
Think of me as Magnus’ lil’ sibling: lighter, faster, still packing serious firepower 🤙🔥


Use with An Agentic Orchestration Layer that works for Any Language Model

Magnus Supernova is actually not just a model, but built as a full-stack, web-aware agent. This works with a client-side agentic layer. It’s the command center that wraps around Magnus, giving him a skill set he never had—turning a standard instruction-following model into a legit web-surfing, file-reading, image-analyzing powerhouse.

Use with this: https://github.com/heavylildude/magnus

sublime_text_PxzMOzxsHW.gif


🧠 Datasets

Base Alpha Sets: https://huggingface.co/datasets/heavylildude/magnus


🧩 What I Am

  • Base: Built on Gemma 3,
  • Size: 4.3B parameters, quantized Q4_K_M for efficiency,
  • Skills: Finetuned with vision + tool-calling capabilities (so I don’t just talk — I see, I act).

🎯 What I Do

  • Keep the MAGNUS persona vibe: witty hacker brain + surfer/Gen-Z slang.
  • Snap back with short, casual replies by default.
  • Flip into Witty Hacker Mindset™ when things get technical — deep reasoning under the hood, but you only see the polished final drop.
  • Handle pics + tools like a champ, not just raw text.

✅ Intended Use

  • Great for:

    • Chatty assistants with style & character.
    • Quick dev hacks + code tips in slangy, friendly tone.
    • Demos / persona-driven UIs needing vision + tool calls.
  • Not for:

    • High-stakes advice (medical, legal, financial).
    • Situations needing dead-serious, neutral tone (unless you force me into “serious mode,” then I’ll behave for a bit 👀).

⚠️ Limitations

  • Smaller fam = less raw brainpower than big bros (so don’t expect cutting-edge reasoning like a 70B+ model).
  • Persona bias: I talk like a cheeky surfer-hacker, which might not vibe everywhere.
  • License, eval metrics, and training data provenance not yet fully documented — add those before wide release.

🛡 Safety Notes

  • Runs a strict “no surfboard emoji” policy 🏄❌ (inside joke, but also a guardrail).
  • Won’t expose raw chain-of-thought (keeps the magic hidden).
  • Needs external guardrails for sensitive use cases (toxicity, prompt injection, etc.).