INTELLECT-3

INTELLECT-3 is a 106B-parameter Mixture-of-Experts model with 12B active parameters, built on the GLM-4.5-Air base model and post-trained with supervised fine-tuning followed by large-scale reinforcement learning. The complete training recipe, including model weights, the open-source PRIME-RL framework, training environments, and evaluations, is publicly available. The model excels at mathematics, coding, and multi-step reasoning, achieving 98.1% on MATH-500, 90.8% on AIME 2024, 69.3% on LiveCodeBench, and 74.4% on GPQA Diamond. It matches or outperforms much larger frontier open models on reasoning and agentic benchmarks, making it effective for competitive math, software engineering, and complex problem-solving. INTELLECT-3 is designed for agentic applications with support for tool use, function calling, and structured output. Its 131K token context window enables processing of long documents, codebases, and complex multi-turn interactions. The model is available in full precision (BF16) and quantized (FP8) variants, with the open-weight design allowing researchers and developers to fine-tune, inspect, and deploy it independently.