nvidia/Nemotron-Research-GooseReason-4B-Instruct
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Jan 14, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Warm

GooseReason-4B-Instruct is a 4 billion parameter reasoning model developed by NVIDIA, fine-tuned from Qwen3-4B-Instruct using Reinforcement Learning with Verifiable Rewards (RLVR) and the Golden Goose pipeline. This model excels in mathematics, programming, STEM reasoning, and logical puzzles, achieving state-of-the-art results among 4B-Instruct models across 15 diverse benchmarks. Its primary differentiator is the use of the GooseReason-0.7M dataset, synthesized from reasoning-rich but previously unverifiable internet text, enabling significant performance gains in complex reasoning tasks.

Loading preview...