Nex-N2-Pro on SiliconFlow: Free Agentic Model Beats Opus 4.7 on Terminal Execution

A new open-weight agentic model from nex-agi just landed on SiliconFlow, and for any operator running terminal-heavy or long-horizon coding workflows, it deserves a routing slot. Nex-N2-Pro scores 75.3 on Terminal-Bench 2.1 — ahead of Claude Opus 4.7 (69.7), DeepSeek-V4-Pro (72.0), and GLM-5.1 (58.7) — while currently pricing at $0.00/M tokens during early access. That combination doesn't stay on the shelf long.

Update, 2026-10-10: Nex-N2-Pro has since been retired by its provider and is no longer served on TheRouter; requests for nex-agi/nex-n2-pro return a deprecation error. The closest live option is qwen/qwen3.5-397b-a17b, the Qwen3.5 base model N2 Pro was post-trained from (same 397B-A17B class, different model). The rest of this article is kept as published.

What Nex-N2-Pro Is

Nex-N2-Pro is the flagship of nex-agi's N2 model family, post-trained on Qwen3.5-397B-A17B (397B total parameters, 17B activated — a MoE architecture). It ships with 262K total context and 256K max output, under the Apache-2.0 license, and is available now on SiliconFlow via serverless inference.

The architecture centers on what nex-agi calls Agentic Thinking: a unified loop that connects task planning, code implementation, environment feedback, evaluation, and iterative debugging into one continuous pass rather than stitching together three separate modules. The practical consequence is that the model keeps driving multi-step terminal or coding tasks forward without losing the thread — exactly the failure mode that makes most "agentic" models frustrating in practice.

Two sub-mechanisms underpin the framework:

• Adaptive Thinking — the model decides autonomously when to reason deeply and when to act quickly. On trivial steps it skips the extended chain-of-thought; on critical decision points it reasons thoroughly. The claimed result is 30–50% fewer thinking tokens versus always-on reasoning, with equal or better performance.
• Coherent Thinking — a single reasoning paradigm is applied consistently across general reasoning, code execution, and browser-based tasks. Capabilities transfer cleanly across modalities instead of fragmenting per task type.

Benchmark Breakdown

The Terminal-Bench 2.1 lead is the signal most relevant to routing teams. Terminal execution is the proxy for agentic loops that touch real environment state: file systems, shell commands, build pipelines, test runners. A 5.6-point margin over Opus 4.7 on that benchmark is material, not noise.

SWE-Bench Pro at 58.8 — edging GPT-5.5 by 0.2 points — confirms the SE capability is genuine, not benchmark-gamed. BrowseComp at 83.7 is within 0.7 points of GPT-5.5 and 0.3 points ahead of DeepSeek-V4-Pro, which puts it squarely in the top tier for deep research tasks.

API Integration

SiliconFlow exposes Nex-N2-Pro through an OpenAI-compatible API endpoint:

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "nex-agi/Nex-N2-Pro",
    "messages": [{"role": "user", "content": "Your prompt here"}],
    "stream": True,
    "max_tokens": 4095,
    "temperature": 0.7,
    "top_p": 0.7,
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json",
}

response = requests.post(url, json=payload, headers=headers)

The model string is nex-agi/Nex-N2-Pro. For operators already routing through SiliconFlow, this is a single-line model ID change. The endpoint is also available through OpenRouter's SiliconFlow provider, making it accessible without a separate SiliconFlow account.

For Claude Code and Anthropic-compatible harnesses, SiliconFlow exposes an Anthropic-format endpoint as well, meaning you can point Claude Code at SiliconFlow and swap in Nex-N2-Pro without touching your toolchain.

Routing Decision: When to Use Nex-N2-Pro

Strong fit:

• Terminal-heavy agentic loops — shell command execution, build pipeline automation, test runner integration. The Terminal-Bench 2.1 lead translates directly here.
• Long-horizon coding tasks — the Agentic Thinking framework is explicitly designed for implement-run-debug cycles that span many steps. SWE-Bench Pro at 58.8 confirms it can drive SE tasks end to end.
• Deep research pipelines — BrowseComp at 83.7 puts it ahead of Opus 4.7 for multi-hop web research and evidence synthesis.
• Cost-sensitive workloads — $0.00/M tokens during early access is zero marginal cost. Run it where you'd otherwise route to a paid tier.

Watch out for:

• Pricing uncertainty — free early access pricing will change. Lock in your routing logic with a price threshold so you don't get surprised when billing starts.
• Context headroom — 262K context is generous but not unlimited. For the very longest transcript-heavy sessions, factor in where truncation happens.
• Provider concentration — Nex-N2-Pro is currently only on SiliconFlow and (via their provider) OpenRouter. If your failover strategy requires provider diversity, you'll need a second fallback.
• Benchmark provenance — the benchmark suite is self-reported by nex-agi. Terminal-Bench 2.1 in particular is relatively new; calibrate against your own eval set before making it your primary terminal agent.

Suggested Routing Policy

# Route to Nex-N2-Pro when:
# - task_type is terminal_execution OR long_horizon_coding
# - cost_tier is free OR budget_sensitive
# - provider: siliconflow OR openrouter/siliconflow
#
# Fallback chain:
# - Primary: nex-agi/Nex-N2-Pro (SiliconFlow)
# - Fallback 1: deepseek-ai/DeepSeek-V4-Pro (when SiliconFlow unavailable)
# - Fallback 2: claude-opus-4-7 (for tasks requiring Anthropic API features)

The free-during-early-access window is the right time to run your own terminal and agentic evals against Nex-N2-Pro. If the numbers hold, you have a strong terminal-execution model at zero marginal cost with an open-weight license — a rare combination.