Nex-N2-Pro is a 396.8B-parameter mixture-of-experts model from Nex-AGI, post-trained on Qwen3.5-397B-A17B, with roughly 17B parameters active per token. It is built around an Agentic Thinking loop that ties requirement understanding, planning, code implementation, environmental feedback, evaluation, and debugging together, with Adaptive Thinking that decides how deeply to reason on each step. The architecture mixes full attention every fourth layer with gated-delta linear attention elsewhere, includes a 27-layer vision tower for reading images alongside text, and carries one multi-token prediction layer, all under an Apache 2.0 release that allows commercial use, modification, and redistribution without royalties.
The model is aimed squarely at agentic engineering work, including agentic coding, deep research, tool calling, and terminal execution, and pairs that focus with a 262,144-token context for reasoning over large codebases and long documents. In benchmark summaries it posts strong but not state-of-the-art numbers, with an 80.8 on SWE-Bench Verified, 75.3 on Terminal-Bench 2.1, 83.7 on BrowseComp, 1585 Elo on GDPval, 90.7 on GPQA Diamond, and a 71.1 self-reported TAU3-Bench score, placing it in useful day-to-day coding territory while trailing the very top closed and open systems on most rows. Because only the safetensors and GGUF builds are available, local use ranges from roughly 96 GB of memory with the IQ1_S quant up through 512 GB for the Q8_0 build, so it fits workstation-class hardware rather than laptops.
