Nex N2.5 Mini and Pro are open-weight multimodal agent models aimed at coding, browser use, computer control, tool calling, and long-running workflows. Mini is the smaller deployment target; Pro trades much heavier infrastructure for stronger vendor-reported results.
Both are currently available through free OpenRouter variants. “Free” describes the current routed endpoint, not a permanent price or production SLA: free-model limits and provider availability can change.
Mini versus Pro at a glance
| Nex N2.5 Mini | Nex N2.5 Pro | |
|---|---|---|
| OpenRouter model | nex-agi/nex-n2.5-mini:free | nex-agi/nex-n2.5-pro:free |
| Current OpenRouter price | Free, rate limited | Free, rate limited |
| Context | 262,144 tokens | 262,144 tokens |
| Input | Text and images | Text and images |
| Output | Text | Text |
| Reasoning modes | None, adaptive, high | None, adaptive, high |
| Tool calling | Supported | Supported |
| Open weights | Yes, Apache 2.0 | Yes, Apache 2.0 |
| Reference self-hosting setup | 2 × H100 | 8 × H100 |
The self-hosting examples are Nex AGI’s reference launch configurations, not minimum hardware guarantees. Mini’s model card reports 35B total parameters. Do not interpret “Mini” as a laptop-class model.
What Nex N2.5 is built for
Nex AGI positions the family around agents that act and verify results in real environments. The model cards emphasize:
- exploring and changing codebases;
- running commands and testing software;
- browser and desktop interaction through visual feedback;
- coding, research, and tool-driven workflows;
- self-correction when observed behavior differs from the intended result.
That positioning is more specific than ordinary chat. It also raises the testing bar: benchmark scores published by Nex AGI are vendor results, and computer-use performance depends on the surrounding harness, tools, permissions, and environment.
When Mini is enough
Choose Mini first when you are prototyping agent loops, testing tool schemas, or comparing models behind an AI gateway. Its free OpenRouter route lowers the cost of evaluation, while its smaller open weights make self-hosting less demanding than Pro—although still far beyond typical consumer hardware in the documented configuration.
Mini is the pragmatic choice when:
- you need agentic coding or browser experiments rather than maximum quality;
- latency and capacity matter more than the last increment of benchmark performance;
- you can tolerate a rate-limited free endpoint;
- you want to validate prompts, tools, and permissions before paying for larger infrastructure.
When Pro makes sense
Pro is for workflows where Mini fails your evaluation set and the quality difference justifies greater serving cost. Nex AGI reports materially higher Pro scores across coding, browser, computer-use, and agent benchmarks, but those results should be reproduced on your own tasks.
Use Pro when multi-step reliability, visual grounding, or difficult software changes matter enough to warrant an eight-H100 reference deployment—or when the free hosted endpoint is adequate for evaluation.
OpenRouter setup
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_API_KEY",
)
response = client.chat.completions.create(
model="nex-agi/nex-n2.5-mini:free",
messages=[{"role": "user", "content": "Inspect this plan and return a safe implementation checklist."}],
tools=[{
"type": "function",
"function": {
"name": "run_check",
"description": "Run an approved project check",
"parameters": {
"type": "object",
"properties": {"command": {"type": "string"}},
"required": ["command"]
}
}
}],
)Switch the model slug to nex-agi/nex-n2.5-pro:free for Pro. Free models are generally limited to 50 requests per day without purchased credits, or 1,000 per day after purchasing at least $10 of credits; verify the live OpenRouter rate-limit policy before relying on it.
Self-hosting and thinking modes
Nex AGI documents SGLang deployment with its own container and Qwen-compatible reasoning and tool parsers. Mini and Pro use --reasoning-parser qwen3 and --tool-call-parser qwen3_coder. The chat template accepts reasoning_effort values for no thinking, adaptive thinking, or always-on thinking.
Self-hosting gives you weight control and predictable routing, but it does not make the models cheap. Account for GPU memory, image inputs, KV cache, 262K-context requests, concurrency, and the customized serving stack. For general deployment planning, use our AI deployment hub and self-hosted versus cloud agent guide.
Limitations
- Current OpenRouter access is free and rate limited; it is not a durable price commitment.
- Nex AGI’s performance tables mix public leaderboard figures and provider evaluations.
- A 262K maximum context does not mean every request should use 262K tokens.
- Computer-use quality depends on the agent harness and permission boundaries.
- Open weights do not remove the need to review training, security, and downstream service terms.
Evaluate both models with a fixed task set, tool-call validation, failure traces, and human review. Our AI Testing & Evaluation hub and agent reliability guide provide that framework.