bolajiev/maxx-1-1.5B

TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 30, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The bolajiev/maxx-1-1.5B is a 1.5 billion parameter Qwen2.5-based instruction-tuned language model developed by bolajiev. Optimized for agentic tasks and instruction following, it is designed for efficient real-world offline use on devices like phones and laptops. This model excels at on-device AI assistant functions, supporting tasks such as summarization, email writing, and multi-step reasoning with a 2048-token context window.

Loading preview...

maxx-1-1.5B: On-Device Agentic LLM

bolajiev/maxx-1-1.5B is a 1.5 billion parameter instruction-tuned model based on Qwen2.5-1.5B-Instruct, specifically optimized for agentic tasks and offline use on personal devices. Developed by bolajiev, this model represents the first checkpoint in a research project aiming to create the best open-source agentic model under 3 billion parameters.

Key Capabilities & Features

  • On-Device Performance: Designed for efficient execution on phones and laptops without internet connectivity.
  • Agentic Task Optimization: Fine-tuned for instruction following, multi-step reasoning, and tool use.
  • Strong Instruction Following: Achieves competitive performance on benchmarks, including a 59.87% MMLU score, outperforming published references for similar-sized models.
  • Privacy-First: All computations run locally on the device.
  • Compact Size: At 1.5B parameters, it's suitable for resource-constrained environments.

Benchmark Highlights (EXP-001)

Evaluated with a 5-shot prompting strategy, maxx-1-1.5B demonstrates strong performance:

  • MMLU: 59.87%
  • TruthfulQA: 45.99% (beating SmolLM2-1.7B by 6 points)
  • Average: 57.75%, within 0.5% of larger competitors like Qwen2.5-1.5B-Instruct and SmolLM2-1.7B-Instruct.

Training Details

The model was fine-tuned using QLoRA (4-bit, rank 16) on a curated dataset of approximately 35,000 examples, including OpenHermes-2.5, UltraChat-200k, Glaive Function Calling v2, and Alpaca Cleaned. Synthetic data was also generated using Qwen2.5-7B as a teacher model. The training was a small run (200 steps) completed in about 1.5 hours on a Kaggle T4 GPU.

Intended Use Cases

  • On-device AI assistants for offline task completion.
  • Summarization, email drafting, scheduling, and planning.
  • Agentic multi-step reasoning for everyday tasks.

Limitations

As an early experimental checkpoint (EXP-001), limitations include a small training run, limited safety alignment, and a context window of 2048 tokens. It has not yet been evaluated on coding tasks.