Nex-N2-Pro open-source model claims top-3 coding benchmark score
Nex AGI's Nex-N2-Pro, a free open-source MoE model with 397B total and 17B active parameters built on Qwen3.5, scores 75.3 on Terminal-Bench 2.1 — placing it top-3 globally among both open and closed models — and introduces "Adaptive Thinking" to dynamically scale reasoning depth per task.
Score breakdown
Evaluate Nex-N2-Pro as a drop-in for agentic coding pipelines — its top-3 Terminal-Bench 2.1 score, 262K context window, and free OpenRouter availability make it a credible open-source alternative to frontier closed models for multi-file refactoring, debugging loops, and chained tool-calling workflows.
- 01Nex-N2-Pro is a free, open-source MoE model by Nex AGI, released June 2, 2026, post-trained on Qwen3.5-397B-A17B.
- 02It has 397B total parameters but only 17B active per forward pass, keeping inference costs low.
- 03Scores 75.3 on Terminal-Bench 2.1 — top-3 globally among both open and closed models.
Nex AGI's Nex-N2-Pro is a free, open-source agentic reasoning model released June 2, 2026, post-trained on Qwen3.5-397B-A17B. As a Mixture-of-Experts (MoE) architecture, it carries 397B total parameters but activates only 17B per forward pass, keeping inference costs closer to a 17B dense model. It ships with a 262,144-token context window, a 256K-token max output, and native support for vision, reasoning, tool calling, structured outputs, and function calling. A smaller companion variant, Nex-N2-mini (35B total, 3B active), targets low-latency and edge deployment scenarios.
This contrasts with conventional reasoning models that apply a fixed chain-of-thought depth regardless of task complexity.
The model's central innovation is "Adaptive Thinking," which the post describes as architecturally enforced at training: simple prompts execute with minimal reasoning overhead, while multi-step agentic tasks trigger deeper, structured planning. This contrasts with conventional reasoning models that apply a fixed chain-of-thought depth regardless of task complexity. Nex-N2-Pro also implements a full agentic coding loop — requirement understanding, task planning, code implementation, environment feedback from actual execution, evaluation, debugging, and continuous iteration — rather than simulating execution.
On benchmarks, the model scores 75.3 on Terminal-Bench 2.1 (top-3 globally among open and closed models), 1585 on GDPval (described as competing with GPT-5.5), and shows strong results on SWE-Atlas and DeepSWE. The post highlights practical strengths in multi-file refactoring, debugging loops, chained tool-calling workflows, and research-to-code tasks. Noted limitations include latency on very large model sizes (the mini variant is suggested for speed-sensitive cases) and gaps in highly specialized domain knowledge. The model is accessible via OpenRouter, SiliconFlow, and self-hosted via a customized `sglang` fork (`nexagi/sglang:v0.5.12`) with weights on HuggingFace at `NexAGI/Nex-N2-Pro`.
Key facts
- 01Nex-N2-Pro is a free, open-source MoE model by Nex AGI, released June 2, 2026, post-trained on Qwen3.5-397B-A17B.
- 02It has 397B total parameters but only 17B active per forward pass, keeping inference costs low.
- 03Scores 75.3 on Terminal-Bench 2.1 — top-3 globally among both open and closed models.
- 04Scores 1585 on GDPval, described as competing with GPT-5.5.
- 05"Adaptive Thinking" dynamically scales reasoning depth per task, architecturally enforced at training.
- 06Context window is 262,144 tokens (262K) with a max output of 256K tokens.
- 07Available for free on OpenRouter; self-hostable via `nexagi/sglang:v0.5.12` with weights at `NexAGI/Nex-N2-Pro` on HuggingFace.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 09:19 UTC. How this works →