INTELLECT-3

INTELLECT-3 - Mixture-of-Experts reasoning model

INTELLECT-3 is presented by Prime Intellect as a 106B-parameter Mixture-of-Experts (MoE) reasoning model (12B active per forward pass) that was post-trained from the GLM-4.5-Air base using supervised fine-tuning followed by large-scale reinforcement learning on Prime Intellect's RL stack.

Defining Characteristics

• Mixture-of-Experts architecture with 106B total parameters and ~12B active per forward pass, documented as the model's structural identity.
• Post-training workflow: SFT on GLM-4.5-Air base followed by large-scale RL using the PRIME-RL trainer.
• Open-source release: weights, training frameworks, environments, and recipes are published under permissive licenses.

Context Strategy

• The model is published with a large context window (documented as 131K tokens for published inference builds) to support long-context tasks.

Evaluation & Positioning

• Prime Intellect reports state-of-the-art performance for the model's size on math, code, science, and multi-step reasoning benchmarks; full benchmark tables and the technical report are published by the authors.