bolajiev/maxx-merged
bolajiev/maxx-merged is a 1.5 billion parameter Qwen2.5-1.5B-Instruct model, fine-tuned by bolajiev for on-device agentic tasks and instruction following. Optimized for offline use on phones and laptops, it focuses on privacy-first AI by performing all computations locally. This model excels at everyday tasks like summarization, email writing, and planning, demonstrating strong commonsense and knowledge retention from its Qwen2.5 foundation.
Loading preview...
maxx-merged: On-Device Agentic LLM (1.5B)
bolajiev/maxx-merged is a 1.5 billion parameter language model, fine-tuned from Qwen2.5-1.5B-Instruct by independent researcher bolajiev. This model is specifically optimized for on-device agentic tasks and instruction following, designed to run efficiently on phones and laptops without requiring internet access. It represents the first checkpoint (EXP-001) in an ongoing research project aiming to develop the best open-source agentic model under 3 billion parameters.
Key Capabilities
- On-Device Inference: Designed for local execution, enabling privacy-first AI applications.
- Instruction Following: Excels at understanding and executing user instructions for various tasks.
- Agentic Reasoning: Capable of multi-step reasoning for everyday tasks like scheduling and planning.
- Strong Foundation: Retains robust commonsense and knowledge base from its Qwen2.5-1.5B-Instruct base.
- Competitive Performance: Achieves an MMLU score of 59.87% and performs within 0.5% of larger competitors on initial benchmarks, notably beating SmolLM2-1.7B on TruthfulQA by 6 points.
Intended Use Cases
- Offline AI Assistant: Ideal for personal assistants that operate without internet connectivity.
- Productivity Tools: Suitable for tasks such as summarization, email drafting, and task management.
- Privacy-Focused Applications: Enables AI functionalities where data remains entirely on the user's device.
This model was trained using QLoRA (4-bit, rank 16) with Unsloth + TRL, on a curated dataset including OpenHermes-2.5, UltraChat-200k, Glaive Function Calling v2, and Alpaca Cleaned data. The current checkpoint has a context window of 2048 tokens.