Spaces:
Sleeping
Apply for a GPU community grant: Personal project
Genesis Manthan-1.5B is a small language model fine-tuned to reason through
tool calls instead of verbal chain-of-thought — trained on Qwen2.5-1.5B-Instruct
using QLoRA SFT + GRPO with tool-execution rewards. It is the first open model
under 2B parameters implementing tool-mediated reasoning (as described in
arXiv:2507.05065), optimized for HuggingFace smolagents workflows.
The demo Space shows users the model's unique behavior: structured
blocks before final answers, rather than verbose verbal reasoning. This visual
proof is the primary way to communicate what makes Manthan different from
standard instruction-tuned models.
The model is 2B parameters (BF16 safetensors). On CPU, inference takes 60-90
seconds per response, making the demo unusable for community evaluation. A T4
GPU would bring this to 3-5 seconds, enabling real interactive use.
Model: https://huggingface.co/Shahansha/Manthan-1.5B
Dataset: https://huggingface.co/datasets/Shahansha/manthan-tool-reasoning-v1
Author: Independent AI researcher, Director of Innovation at Arcadis,
founder of Genesis AGI (https://genesisagi.in) — running free community
AI education sessions in India.