Nex-AGI launches Nex-N2.5 agentic model family (mini, Pro, 1.6T-parameter MoE Max) with open weights and OpenRouter access

Nex-AGI's Nex-N2.5 family spans mini, Pro, and a 1.6-trillion-parameter text-only MoE Max built for long-horizon agentic tasks; Max scores 86.1 on Terminal-Bench 2.1 and 65.7 on SWE-Bench Pro (trailing Claude Opus 5), while Pro scores 82.7 and 61.2; weights…

Two Hugging Face trending-model reports on 2026-09-08 (05:08 UTC and 11:01 UTC, each showing a Nex-N2.5 variant at #30) cover the launch of Nex-N2.5, a next-generation family of agentic models in three sizes: mini, Pro, and Max. Nex-N2.5-Max is built on a 1.6-trillion-parameter text-only Mixture-of-Experts foundation, marking the company's first complete post-training effort at trillion-parameter scale. The models target long-horizon agentic tasks, including computer use, web browsing, autonomous program execution, and visually grounded agentic capabilities with self-correction, supported by expanded agent training environments. Reported benchmarks: Max scores 86.1 on Terminal-Bench 2.1 and 65.7 on SWE-Bench Pro, trailing Claude Opus 5; Pro scores 82.7 on Terminal-Bench 2.1 and 61.2 on SWE-Bench Pro. The later report describes comparisons against Claude Opus 5, GPT-5.6 Sol, Kimi-K3, GLM-5.3, DeepSeek-V4-Pro-0813, and Qwen3.8-Max as showing competitive scores on both benchmarks. Distribution is via open weights on Hugging Face and ModelScope with hosted access through OpenRouter; the earlier report said weights 'are being released openly,' while the later report said weights 'will be released' and were listed as 'coming soon' at publication, so the exact release status is disputed between the two snapshots.

• Nex-N2.5 family has three sizes: Nex-N2.5-mini, Nex-N2.5-Pro, and Nex-N2.5-Max; mini trended #30 on Hugging Face at 05:08 UTC and Pro trended #30 at 11:01 UTC on 2026-09-08.
• Nex-N2.5-Max is built on a 1.6-trillion-parameter text-only Mixture-of-Experts foundation — Nex-AGI's first complete post-training effort at trillion-parameter scale.
• Max benchmarks: 86.1 on Terminal-Bench 2.1 and 65.7 on SWE-Bench Pro, trailing Claude Opus 5.
• Pro benchmarks: 82.7 on Terminal-Bench 2.1 and 61.2 on SWE-Bench Pro.
• The later report lists comparisons against Claude Opus 5, GPT-5.6 Sol, Kimi-K3, GLM-5.3, DeepSeek-V4-Pro-0813, and Qwen3.8-Max, described as competitive on Terminal-Bench 2.1 and SWE-Bench Pro.
• Target capabilities: long-horizon computer use, web browsing, autonomous program execution, and visually grounded agentic tasks/self-correction, with expanded agent training environments.
• Distribution: open weights on Hugging Face and ModelScope; hosted access via OpenRouter.
• Sources disagree on release status: the 05:08 UTC report said weights 'are being released openly'; the 11:01 UTC report said weights 'will be released' and were listed as 'coming soon' at publication.