nex-agi/Nex-N2.5-mini

TEXT GENERATIONPricing:Input $0.1 / Cached $0.07 / Output $1Concurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 8, 2026License:apache-2.0Architecture:Transformer0.7K Open Weights Featherless Exclusive Cold

Nex-N2.5-mini is a 35.1 billion parameter agentic model developed by Nex-AGI, designed for long-horizon tasks in real-world environments. It builds on the multimodal foundations of Nex-N2, with focused improvements in computer use, web browsing, and visually grounded agentic capabilities. This model excels at continuous action and self-correction through visual feedback, enabling it to operate computers and browsers, and autonomously execute and test programs. It is optimized for scientific research, knowledge work, and complex productivity tasks.

Loading preview...

Nex-N2.5-mini: Agentic Model for Real-World Tasks

Nex-N2.5-mini, developed by Nex-AGI, is part of a new family of agentic models specifically engineered for tackling long-horizon tasks within real-world environments. This 35.1 billion parameter model enhances the multimodal capabilities of its predecessor, Nex-N2, with significant advancements in computer and web interaction, and visually grounded agentic functions.

Key Capabilities

  • Visually Grounded Agentic Behavior: The model leverages visual feedback as a critical interface for perceiving environments, verifying outcomes, and progressing tasks, enabling continuous action and self-correction.
  • Autonomous Computer and Browser Operation: Nex-N2.5-mini can operate computers and browsers, and autonomously execute and test programs.
  • Enhanced Agent Training: It benefits from expanded agent training environments and task types, leading to improved performance in complex scenarios.
  • Robust Function Calling: Supports function calling, configurable via the --tool-call-parser qwen3_coder flag.
  • Adaptive Thinking Modes: Features reasoning_effort settings ("none", "medium", "high") to control the model's thinking behavior, allowing it to adaptively decide on the extent of pre-response reasoning.

Performance Highlights

Nex-N2.5-mini demonstrates strong performance across various benchmarks, particularly in agentic workflows, computer use, and multimodal understanding. For instance, it achieves 73.4 on Terminal-Bench 2.1, 54.6 on Toolathlon Verified, and 83.4 on BrowseComp. In multimodal tasks, it scores 71.2 on OSWorld-Verified and 82.9 on OSWorld-G.

Good For

  • Scientific Research: Automating complex research workflows.
  • Knowledge Work: Handling intricate information processing and analysis tasks.
  • Complex Productivity: Streamlining multi-step tasks requiring interaction with digital environments.
  • Developers: Integrating agentic capabilities into applications, with open-source weights and Docker deployment options.

Finetunes
2 models