intellect-3 is here. this model has been my life the last 2 months and fucked my sleep schedule more than once. a huge team effort that im very proud of. some technical details:
we train on top of glm-4.5-air-base, a 106b param moe, and do the entire post-training pipeline in-house on our open-source stack prime-rl+verifiers+hub. this means two sft stages, a general reasoning and agentic phase, followed by a large-scale rl training run
