INTELLECT-3

### TL;DR

A 100B+ parameter Mixture-of-Experts language model trained with large-scale reinforcement learning, achieving state-of-the-art performance on math and reasoning benchmarks. Built on GLM-4.5-Air base with 131K context window.

Key Insights & Metrics

Key Features

• 106B parameter Mixture-of-Experts architecture trained with Prime Intellect RL stack
• State-of-the-art math and reasoning performance — evaluated via PRIME-RL and verifiers
• $0.20 input / $1.10 output per 1M tokens; 131K token context window with tool use support

→ Related Releases

Discussion