Nex-N2 open-sources agentic models that think before they act

Nex-N2 is a new family of open-source agent reasoning models from nex-agi, purpose-built for real-world coding and long-horizon autonomous tasks. Instead of treating reasoning, tool calling, and environment execution as isolated steps, it ties them into a single loop where the model actually knows when to pause and think, and when to just execute.

Post-trained on Alibaba's Qwen 3.5 series, the release includes two Mixture-of-Experts variants. The flagship Nex-N2-Pro packs 397 billion total parameters but only activates 17 billion per token. The lightweight Nex-N2-mini sits at 35 billion parameters with just 3 billion active. Keeping active parameters lean is key here, making multi-step agentic workflows far more viable on practical hardware.

What it does

Most reasoning models suffer from overthinking, burning expensive tokens and time on brainless tasks like listing a directory. Nex-N2 tackles this through adaptive thinking. For straightforward terminal commands, it acts immediately. When it hits ambiguous requirements, tricky architecture decisions, or complex debugging loops, it spins up deep reasoning on demand.

Nex-agi frames this around an Agentic Thinking framework that unifies requirement parsing, planning, coding, environment feedback, and testing into one continuous cycle. Rather than switching personas between searching and coding, the model maintains a coherent reasoning pattern across modalities and tool calls.

In real setups, the team tested it across one-person company agent workflows, end-to-end game prototyping, and terminal automation. Benchmark-wise, Nex-N2-Pro posted 75.3 on Terminal-Bench 2.1, 1585 on GDPval, and 33.6 on DeepSuite, pulling ahead of prior open weights like DeepSeek V4 and GLM 5.1, while rivaling frontier closed models like GPT-5.5 and Opus 4.7 across several software engineering suites.

How to run it

Both Nex-N2-Pro and Nex-N2-mini weights are available now on Hugging Face and ModelScope. If you want to test them without spinning up servers, early API access is live on SiliconFlow.

For self-hosting, the team recommends their customized sglang fork, which includes specific parsers for Qwen 3 tool calling and reasoning tokens. Running the mini model is relatively approachable, needing just a single node with 2× H100 GPUs. The Pro model demands real enterprise muscle, with the official launch script configured for two nodes of 8× H100s across 16-way tensor parallelism. A prebuilt Docker image is also provided to skip dependency wrangling.

Why it matters

If you build coding agents or autonomous terminal bots, Nex-N2 is worth testing right now. Open models usually fall apart during step 15 of a 30-step task, getting stuck in repetitive loops when bash commands fail. Nex-N2's unified loop and explicit focus on environment feedback directly target that breaking point.

The clever move here is pairing dynamic reasoning with a sparse MoE architecture. For builders running automated pipelines, paying full inference compute on every trivial shell command destroys margins. Pushing active parameters down to 17B and 3B makes agent loops affordable without lobotomizing decision quality. Just keep in mind that agentic power is only as good as the sandbox you build around it; your harness engineering still determines whether it flies or crashes.