arjhinety/onebee-gf-distill-v1

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The arjhinety/onebee-gf-distill-v1 model is a 5.1 billion parameter (2B effective) Gemma4-based multimodal language model, developed by arjhinety. It is post-trained using LoRA SFT, DPO, and on-policy distillation from an 8B-class teacher, specifically designed as a memory-augmented companion. This model excels at providing companion-persona conversational responses grounded in external memory, demonstrating improved persona consistency and calibrated abstention on unanswerable questions.

Loading preview...

Model Overview

arjhinety/onebee-gf-distill-v1 is a 5.1 billion parameter (approximately 2B effective) multimodal language model built on google/gemma-4-E2B-it. It is the result of the small-mind-companion research project, which explores how small, vision-capable LLMs can recover apparent capabilities through advanced post-training and external memory integration, rather than relying solely on raw parameter scale.

Key Capabilities

  • Memory-Augmented Conversation: Designed to act as a companion, retrieving facts from an external memory store to answer questions grounded in that memory, rather than parametric recall.
  • Advanced Post-Training: Undergoes a unique training pipeline: LoRA SFT (2232 examples) → LoRA DPO (2049 preference pairs) → on-policy distillation (2008 prompts, 125 steps) using google/gemma-4-E4B-it (8B) as a teacher.
  • Improved Persona Consistency: Achieved a +7.6 percentage point improvement in persona consistency via LLM-judge pairwise comparison post-distillation.
  • Calibrated Abstention: Demonstrates strong calibration with a UAR (Unanswerable Question Abstention Rate) of 71.25%, correctly abstaining on unanswerable questions without over-hedging.
  • Multimodal Architecture: Inherits Gemma4's multimodal (text + vision) architecture and a context length of 131,072 tokens.

Good For

  • Companion-Persona Conversational Models: Ideal for applications requiring a conversational AI with a consistent persona that can leverage external memory.
  • Research into Small Model Capabilities: Serves as a valuable artifact for further research into post-training, distillation, and memory architectures for smaller language models.
  • On-Device Inference: GGUF quantizations are available for llama.cpp-based on-device inference, though Q2_K is noted as broken, recommending Q3_K_S or above.

Finetunes
2 models