madhun/Nex-N2.5-mini
The Nex-N2.5-mini is a 35.1 billion parameter multimodal model developed by Nex-AGI, part of the Nex-N2.5 family of agentic models. It is designed for long-horizon tasks in real-world environments, with focused improvements in computer use, web browsing, and visually grounded agentic capabilities. This model strengthens its ability to act continuously and self-correct through visual feedback, enabling it to operate computers and browsers autonomously. It excels in agentic workflows, coding, and multimodal understanding, supporting a context length of 32768 tokens.
Loading preview...
Nex-N2.5-mini: An Agentic Multimodal Model for Real-World Tasks
Nex-N2.5-mini, developed by Nex-AGI, is a 35.1 billion parameter model within the Nex-N2.5 family, specifically engineered for agentic capabilities in complex, real-world environments. It builds upon the multimodal foundations of its predecessor, Nex-N2, with significant enhancements in computer and web interaction.
Key Capabilities
- Agentic Intelligence: Designed for long-horizon tasks, enabling continuous action and self-correction through visual feedback.
- Multimodal Understanding: Integrates vision as a critical interface for perceiving environments, verifying outcomes, and advancing tasks.
- Computer and Browser Operation: Can autonomously operate computers and browsers, execute, and test programs.
- Coding Proficiency: Demonstrates strong performance across various coding benchmarks, including Terminal-Bench 2.1 (73.4), SWE-Bench Pro (43.8), and DeepSWE v1.1 (36.1).
- Agentic Workflow Benchmarks: Achieves notable scores in AutomationBench v1.0.6 (32.3), Toolathlon Verified (54.6), and BrowseComp (83.4).
- Multimodal Benchmarks: Performs well in OSWorld-Verified (71.2), WebTest (48.6), and Vision2Web (52.9).
- Configurable Reasoning: Supports different "thinking modes" (none, medium, high) via
reasoning_effortto control the model's deliberation before responding. - Function Calling: Equipped with robust function-calling capabilities, enabled by the
--tool-call-parser qwen3_coderflag.
Good For
- Automated Task Execution: Ideal for scenarios requiring autonomous interaction with digital environments, such as web navigation or software testing.
- Complex Problem Solving: Suitable for tasks that benefit from an agentic approach, where continuous feedback and self-correction are crucial.
- Developer Tools: Can be integrated into systems requiring advanced code generation, debugging, or automated development workflows.
- Research and Productivity: Applicable for scientific research, knowledge work, and other complex productivity tasks that leverage its multimodal and agentic strengths.