Moonshot AI has introduced the Nex-N2 model family, positioning it as a system designed not just to reason, but to act. The launch centers on Nex-N2-Pro and smaller variants that the company says are tuned for real-world productivity, coding, search and agent-style tool use.
The Chinese AI lab describes the family with the slogan “thinking, built for action.” Rather than treating reasoning and execution as separate behaviors, Moonshot AI says Nex-N2 uses a unified approach across search, coding and tool calling. The company says the model follows a consistent internal process that includes breaking down goals, tracking state, adjusting strategy and checking its own work. That approach, it argues, helps especially in mixed tasks that require switching between reading, writing code and calling tools.
Moonshot AI is highlighting benchmark results to support the launch. In its materials, Nex-N2-Pro is shown against several prominent models on productivity and software-focused tests. On Terminal-Bench 2.1, the company lists Nex-N2-Pro at 75.3, ahead of Claude Opus 4.7 at 69.7 and DeepSeek V4 Pro at 72.0, though behind GPT-5.5 at 83.4. On SWE-bench Pro, Nex-N2-Pro is listed at 58.8, close to GPT-5.5 at 58.6 and DeepSeek V4 Pro at 55.4, but below Claude Opus 4.7 at 64.3.
The company also cites results on DeepSWE, GDPval, BrowseComp and Toolathlon. Across those tests, Nex-N2-Pro appears competitive in some areas and behind leading rivals in others. The overall message from Moonshot AI is that the model is intended to keep pace with top-tier systems on long-running, action-heavy workloads rather than just short-form chat.
A major theme in the launch is what Moonshot AI calls adaptive thinking. According to the company, Nex-N2 can decide when to activate more intensive reasoning and when to keep processing lighter. Moonshot AI says this dynamic approach is meant to preserve task performance while reducing unnecessary token use.
The company claims this setup helps the model spend more reasoning effort where uncertainty is highest, such as during bug localization, verification steps or the end stages of a search task. In longer workflows, Moonshot AI says the model concentrates its thinking around key decision points instead of distributing it evenly throughout a task.
Moonshot AI says the smaller Nex-N2-mini model also benefits from the adaptive approach. In its own comparisons, the company says the model performs much better than when chain-of-thought is disabled and roughly matches or slightly exceeds always-on reasoning, while lowering overall token costs by about 20%.
The company says Nex-N2-Pro is available to try through multiple platforms, including OpenRouter, SiliconFlow, Hugging Face, ModelScope and GitHub. That distribution suggests Moonshot AI is aiming for broad access among developers and researchers who want to test the model in coding and agentic workflows.
Moonshot AI has framed the release as an attempt to close the gap between reasoning models and practical assistants that can complete work rather than merely discuss it. Whether Nex-N2 ultimately changes developer preferences will depend on how the models perform outside the company’s own evaluations, but the launch shows continued competition among AI labs to build systems optimized for long-horizon, action-based tasks.