Overview

This model, DJLougen/Nemotron-Research-GooseReason-4B-Instruct-MLX-16bit, is an MLX-optimized, 16-bit full-precision conversion of NVIDIA's Nemotron-Research-GooseReason-4B-Instruct. Built upon the Qwen3-4B-Instruct-2507 base model, it features 4.4 billion parameters and a substantial maximum sequence length of 32,768 tokens. The original model was developed by NVIDIA using Reinforcement Learning with Verifiable Rewards (RLVR) to enhance its reasoning capabilities.

Key Capabilities

• Advanced Reasoning: Specifically optimized for complex reasoning across various domains.
• Math Performance: Demonstrates strong results on benchmarks like AIME 2025 (55.0 avg@64) and AMC (82.2 avg@64).
• Code Generation: Achieves competitive performance on coding benchmarks such as LiveCodeBench v6 (30.1 pass@1) and HumanEval.
• STEM Reasoning: Possesses broad scientific and technical reasoning abilities.
• Thinking Mode: Utilizes an extended thinking mode with <think> tags for tackling intricate problems, which can be explicitly prompted.

Good For

• Mathematical Problem Solving: Ideal for applications requiring high accuracy in math and arithmetic.
• Code Development: Suitable for generating and understanding code snippets.
• Scientific and Technical Analysis: Effective in tasks demanding STEM-related reasoning.
• Complex Reasoning Tasks: Benefits from its RLVR training and thinking mode for multi-step logical deductions.