Open-source 1.6T-param MoE Nex-N2.5-Max lands 0.1 points behind Claude Opus 5 on agent benchmark
airesearch12 · x · 2026-09-09
- NexEcosystem released Nex-N2.5-Max, an open-source 1.6-trillion-parameter MoE built for agents and long-horizon workflows: coding + tool use, 262K context, text-only, Apache 2.0, weights already on Hugging Face.
- Reported results: AutomationBench — Claude Opus 5 → 50.3, Nex-N2.5-Max → 50.2, GPT-5.6 Sol → 45.8, Qwen3.8-Max → 39.8; BrowseComp — Nex 92.6 vs Claude 90.8; Terminal-Bench 2.1 → 86.1; Toolathlon Verified → 74.7. Beats DeepSeek V4 Pro 0813 on all but one comparable benchmark.
- The catch: the repo is 1.65TB and recommended deployment needs 16× H200-class hardware, so local runs are out of reach for individuals.
Related event: Nex-AGI Open-Sources Nex-N2.5 Agentic Model Family up to 1.6T Parameters(5 posts)→
More from coding & agent
- OpenAI rolls out Astra model to all Codex and ChatGPT Work users — craigsdennis · 2026-09-09
- Reading video frame-by-frame in Codex with ffmpeg: a clever workflow tip — _AustinCalvert_ · 2026-09-09
- Gemini CLI v0.60.0-preview hardens sandboxing, OAuth and path security — gemini-cli-robot · 2026-09-09
- W&B's research agent ARIA impresses: custom experiment analysis with auto-generated Plotly charts — _ScottCondron · 2026-09-09
- Harvey post-trains Qwen3.5 orchestrator: rubric pass rate jumps from 29.9% to 63%, beats Claude Code — SergioPaniego · 2026-09-09
- Agents generating Python per action make traces far harder to debug than built-in tool calls — BraceSproul · 2026-09-09