Nex-N2-Pro

Nex-N2-Pro is the larger of the two models Nex AGI open-sourced on 3 June 2026 as its second generation, alongside the smaller Nex-N2-mini. Both are post-trained on the Qwen3.5 series rather than trained from scratch: the Pro edition on Qwen3.5-397B-A17B, the mini on Qwen3.5-35B-A3B-Base. The maker states this openly in the model card, which puts the profile of both models on borrowed foundations and makes Nex AGI's contribution the post-training, not the architecture. The pitch is agentic work rather than conversation. Nex AGI describes an Agentic Thinking framework in two parts: Adaptive Thinking, where the model decides on its own when to reason and how deeply, executing simple actions fast and deliberating on critical ones; and Coherent Thinking, meant to carry one reasoning paradigm across general reasoning and agentic tasks so that capability transfers between them. The stated target is long-horizon work with environmental feedback — coding, deep research, tool calling and terminal execution. The published benchmark table is unusually complete: it reports Nex-N2-Pro and Nex-N2-mini side by side against GPT-5.5, Opus 4.7, Kimi-K2.6, GLM-5.1, MiniMax M3 and DeepSeek-V4-Pro across sixteen tests. On that table the Pro edition scores 75.3 on Terminal-Bench 2.1, 80.8 on SWE-Bench Verified, 58.8 on SWE-Bench Pro, 83.7 on BrowseComp, 90.7 on GPQA Diamond and 1585 on GDPval. Nex AGI does not claim the top of the table and the numbers do not give it: GPT-5.5 and Opus 4.7 win most rows, and on DeepSWE the gap is wide — 33.6 against 70 for GPT-5.5. What the table does show is a consistent second tier, and a mini edition that holds within ten points of its larger sibling on several agentic rows while collapsing on DeepSWE (8.0). Deployment is documented for the maker's own fork of sglang. The Pro edition is launched across two nodes of eight H100 accelerators each with tensor parallelism of 16; the mini runs on a single machine with two H100s. Nex AGI ships a prebuilt Docker image, recommends temperature 0.7 with top-p 0.95 and top-k 40, and supports function calling and separated reasoning traces through dedicated parsers. Weights went out on Hugging Face and ModelScope under Apache 2.0, with early hosted access through SiliconFlow and two free weeks on OpenRouter from 9 June 2026. Nex AGI publishes no context length for this generation in the model card and no price list of its own — it is a weights-first release.

▸Videos

No videos yet.

▸Reviews