Nex-N2.5-mini Review: A Compact Agentic Model for Computer Use and Web Tasks
Evaluating Nex-AGI's smallest agentic model for lightweight workflows, computer use, and web browsing tasks - free, open-source, and capable for experimentation.
Quick Answer Table
| Model | Best For | Price (per 1M tokens) | Pick If |
|---|---|---|---|
| Nex-N2.5-mini | Lightweight agentic tasks, computer use | Free (OpenRouter) | You need a free, open-source model for basic agentic workflows |
| Nex-N2.5-Pro | Balanced performance, multimodal | TBD | You need stronger agentic capabilities with vision |
| Nex-N2.5-Max | Complex reasoning, trillion-parameter scale | TBD | You need maximum performance for research/productivity |
| Claude Opus 5 | Top-tier agentic performance | $5 input / $25 output | You need the best available agentic performance |
| GPT-5.6 Sol | General purpose with strong agentic | $2 input / $10 output | You need a versatile model with good agentic skills |
Overview
Nex-N2.5-mini is the smallest member of the Nex-N2.5 family of agentic models released by Nex-AGI on September 8, 2026. Designed for lightweight agentic workflows, it focuses on computer use, web browsing, and visually grounded tasks while maintaining strong performance in coding benchmarks.
Key Specifications
- Release Date: September 8, 2026
- Parameters: 35.1B total (BF16); Mixture-of-Experts with 256 experts, 8 active per token (config)
- Architecture: Vision-language MoE (
Qwen3_5MoeForConditionalGenerationwith avision_config, per the config); the model card says mini builds on the multimodal foundations of Nex-N2, even though its Hugging Facepipeline_tagistext-generation - License: Apache 2.0
- Context Length: Not specified in README
- Access:
- Hugging Face: https://huggingface.co/nex-agi/Nex-N2.5-mini
- OpenRouter (free tier): https://openrouter.ai/nex-agi/nex-n2.5-mini
- Docker: nexagi/sglang:v0.5.18-nex-patch
Performance Analysis
Coding Benchmarks
Nex-N2.5-mini shows solid coding performance, though it trails the larger variants and top competitors:
- Terminal-Bench 2.1: 73.4% (vs 89.1% for Claude Opus 5)
- SWE-Bench Pro: 43.8% (vs 79.2% for Claude Opus 5)
- DeepSWE v1.1: 36.1% (vs 73.7% for Claude Opus 5)
Agentic Workflows
The model demonstrates capable agentic performance for its size:
- AutomationBench v1.0.6: 32.3% (vs 50.3% for Claude Opus 5)
- Toolathlon Verified: 54.6% (vs 76.5% for Claude Opus 5 and Kimi-K3)
- Job Bench: 28.5% (vs 65.7% for Claude Opus 5)
- BrowseComp: 83.4% (strong web browsing capability)
Computer Use
Strong performance in computer use tasks:
- OSWorld-Verified: 71.2% (vs 86.1% for Qwen3.8-Max)
- WebArena-Verified: 63.4% (vs 71.6% for Kimi-K3)
- Vision2Web: 52.9% (vs 79.8% for GPT-5.6 Sol)
Pricing and Accessibility
- OpenRouter: Available for free usage (rate limits may apply)
- Hugging Face: Weights available under Apache 2.0 license
- Self-hosted: Docker image provided (nexagi/sglang:v0.5.18-nex-patch)
- No official per-token pricing published - free access suggests community/research focus
Use Case Recommendations
Choose Nex-N2.5-mini if:
- You need a free, open-source agentic model
- Your use case involves light computer use or web automation
- You're experimenting with agentic workflows and want zero-cost entry
- You prefer self-hosting with Docker for data privacy
Consider alternatives if:
- You need maximum agentic performance (consider Claude Opus 5 or Nex-N2.5-Max)
- You have high-volume production needs (free tier may have limits)
Verdict
The Nex-N2.5-mini represents a solid entry point into the Nex-N2.5 agentic family, offering capable performance for lightweight agentic tasks at zero cost. While it doesn't match the top-tier performance of Claude Opus 5 or the larger Nex-N2.5 variants, its free availability and strong showing on BrowseComp (83.4%), which the model card lists among its text agentic benchmarks, make it an attractive option for developers exploring agentic workflows without financial commitment. The Apache 2.0 license and Docker deployment option further enhance its appeal for experimentation and privacy-sensitive applications.
Sources
- Model card: https://huggingface.co/nex-agi/Nex-N2.5-mini
- README: https://huggingface.co/nex-agi/Nex-N2.5-mini/resolve/main/README.md
- Benchmark data: Extracted from model card tables
- OpenRouter listing: https://openrouter.ai/nex-agi/nex-n2.5-mini