LoNebula/my-jarvis-72b

TEXT GENERATIONPricing:Input $12 / Output $20Concurrent Unit Cost:4Model Size:72.7BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

LoNebula/my-jarvis-72b is a 72.7 billion parameter bilingual (Japanese/English) personal assistant model fine-tuned from Qwen2.5-72B-Instruct. It is designed as a blank slate, with no pre-loaded personal data or identity, ensuring personalization occurs entirely on the user's device after loading. This model is optimized for local, user-specific adaptation through onboarding questions, conversation memory, and optional per-user LoRA adapters, making it suitable for privacy-focused personal AI assistant applications.

Loading preview...

My-Jarvis 72B Overview

My-Jarvis 72B is a 72.7 billion parameter bilingual (Japanese/English) personal assistant model developed by LoNebula. It is fine-tuned from the Qwen2.5-72B-Instruct base model using 4-bit QLoRA, making it trainable on hardware like a single HAKUSAN GPU node (2 x A40 48GB). A key differentiator is its design philosophy: the released model weights contain no owner or user personal data, ensuring a privacy-first approach where personalization begins only after the user loads the model.

Key Capabilities & Features

  • Bilingual Support: Proficient in both Japanese and English.
  • Privacy-Centric Design: The model is a "blank slate" with no pre-defined persona or personal data embedded in its weights. A PII scanner is used during development to prevent any personal data inclusion.
  • On-Device Personalization: All personalization happens locally on the user's machine, including:
    • Onboarding: Initial questions to establish user name, interests, and tone.
    • Proactive Memory: Distills new user facts from conversations into local memory for context retrieval.
    • Per-User LoRA Adapter (Optional): Users can train small, personalized LoRA adapters locally from their conversation memory, which never leaves their device.
  • Base Model: Built upon the robust Qwen2.5-72B-Instruct, offering strong general language understanding.

Good For

  • Developers building privacy-focused personal AI assistants.
  • Applications requiring a bilingual assistant that adapts to individual user preferences over time.
  • Use cases where user data must remain strictly on the user's device.
  • Experimentation with large language models that offer a clean, unpersonalized starting point.

Finetunes
1 models