microsoft-optimind-sft
Hugging FaceVersion: 5
OptiMind-SFT is a specialized 20B parameter model designed to bridge the gap between natural language and executable optimization solvers. It automates the translation of complex decision-making problems—such as supply chain planning, scheduling, and resource allocation—into correct MILP formulations.
Developer: Microsoft Research, Machine Learning and Optimization (MLO) Group
Model Architecture: Mixture-of-Experts (MoE) variant of the transformer architecture (gpt-oss family).
Parameters: 20 Billion (3.6B activated)
Inputs: Natural language optimization problem description.
Context Length: 128,000 tokens
Outputs: Mathematical formulation and executable Python code using GurobiPy.
GPUs: 8x NVIDIA B200 (Training), 8x NVIDIA H100 (Inference/Evaluation)
Training Time: ~8 hours
Public Data Summary: Cleaned subsets of OR-Instruct and OptMATH-Train
Dates: Trained in October 2025
Status: Static model trained on cleaned public datasets
Release Date: November 2025
License: MIT
Model Dependencies: unsloth/gpt-oss-20b-BF16
Additional Assets: GitHub Repository
Model Architecture: Mixture-of-Experts (MoE) variant of the transformer architecture (gpt-oss family).
Parameters: 20 Billion (3.6B activated)
Inputs: Natural language optimization problem description.
Context Length: 128,000 tokens
Outputs: Mathematical formulation and executable Python code using GurobiPy.
GPUs: 8x NVIDIA B200 (Training), 8x NVIDIA H100 (Inference/Evaluation)
Training Time: ~8 hours
Public Data Summary: Cleaned subsets of OR-Instruct and OptMATH-Train
Dates: Trained in October 2025
Status: Static model trained on cleaned public datasets
Release Date: November 2025
License: MIT
Model Dependencies: unsloth/gpt-oss-20b-BF16
Additional Assets: GitHub Repository