Nex-N2-Pro: The Free Open-Source Model That Rivals GPT-5.5, and Its N2.5 Successor

A Chinese AI lab just released a free model that scores 80.8 on SWE-Bench Verified and 75.3 on Terminal-Bench 2.1, putting it in the same league as GPT-5.5 and Claude Opus 4.7 on real coding work. It is called Nex-N2-Pro, it dropped on June 2, 2026, and it costs nothing to download. The model comes from Nex AGI, ships under a permissive Apache 2.0 license, and is built specifically for agentic tasks like writing code, calling tools, and running long, multi-step workflows on its own.

Five months on, Nex-N2-Pro has a successor. Nex AGI released the Nex-N2.5 family (mini, Pro and Max) on September 7 and 8, 2026, and hosted providers have since dropped N2-Pro, so the API route to Nex AGI’s models now runs through Nex-N2.5-Pro. The launch analysis below still holds for N2-Pro, and the new family has its own section.

This is one of the clearest signs yet that open-source AI has caught up with the frontier on the tasks developers actually care about. In this article we break down what Nex-N2-Pro is, what its benchmark scores really mean, how it stacks up against GPT-5.5 and Opus 4.7, what it costs, and who should actually bother with it. We also cover the part most launch posts skip, which is the very real hardware needed to run a 397-billion-parameter model on your own machine.

A permissive license like Apache 2.0 is what makes Nex-N2-Pro genuine open source AI, free to download, modify, and run on your own hardware.

Nex-N2-Pro also features in our roundup of the best open source AI models as the most permissively licensed agentic pick.

The Key Takeaways

• Nex-N2-Pro is free and open-source under Apache 2.0, released by Nex AGI on June 2, 2026
• It is a 397B-parameter Mixture-of-Experts model that activates only 17B parameters per token, built on Qwen3.5-397B-A17B
• Headline scores include 80.8 SWE-Bench Verified, 75.3 Terminal-Bench 2.1, 90.7 GPQA Diamond, and 94.0 IFEval
• At launch it kept pace with GPT-5.5 and Claude Opus 4.7 on coding and agentic benchmarks, the first open model to do so this convincingly
• A smaller sibling, Nex-N2-mini, targets lighter hardware, but running the full Pro model locally needs roughly 794 GB of VRAM at full precision
• Nex-N2.5 replaced it in September 2026. Hosted N2-Pro is largely gone, and Nex-N2.5-Pro costs $0.075 per million input tokens and $0.25 per million output tokens on OpenRouter

What Is Nex-N2-Pro?

Nex-N2-Pro is an open-source AI model from Nex AGI designed for agentic work rather than simple chat. It uses a Mixture-of-Experts (MoE) design with 397 billion total parameters but activates only 17 billion per token, which keeps inference far cheaper than a dense model of the same size. It accepts both text and images as input and produces text output, and it supports a 262,000-token context window with output up to 256,000 tokens.

The model is post-trained on Qwen3.5-397B-A17B, Alibaba’s open base model, so it inherits that lineage while adding Nex AGI’s own agentic training on top. Nex-N2-Pro is the larger of two releases. Its sibling, Nex-N2-mini, is built on the smaller Qwen3.5-35B-A3B base for people who want something lighter to run. Both are open under Apache 2.0, which means you can use them commercially, modify them, and self-host them without paying a license fee.

If you have followed recent open releases like Alibaba’s Qwen pricing and model lineup or MiniMax M3, Nex-N2-Pro fits the same wave of capable, openly licensed models. The difference is how aggressively it targets agentic coding.

Nex-N2-Pro Benchmarks: How Good Is It Really?

The benchmark sheet is where Nex-N2-Pro earns its attention. On software engineering and agentic tasks, it posts numbers that until recently were the exclusive territory of closed frontier models. The figures below come from the official Hugging Face model card.

The standout is 80.8 on SWE-Bench Verified, a test that asks a model to fix real bugs in real open-source repositories. A score in the 80s means the model is useful for day-to-day engineering work, not just demos. The 94.0 on IFEval also matters for agent builders, because an agent that drifts from instructions mid-task is worse than no agent at all.

How Nex-N2-Pro Compares to GPT-5.5 and Opus 4.7

Nex AGI positions the model as a peer to the closed leaders, and the head-to-head numbers it published back that up on coding. Nex-N2-Pro edges GPT-5.5 on SWE-Bench Pro and beats Claude Opus 4.7 on Terminal-Bench 2.1, while trailing GPT-5.5 slightly on web browsing.

The honest read is that Nex-N2-Pro is competitive, not dominant. It trades blows with the best closed models on coding and agentic work, loses by a hair on some browsing tasks, and wins big on one metric that the others cannot match, which is price. You can read our take on the closed competition in our Claude Opus coverage and see where it sits on our best AI models hub.

What Makes “Agentic Thinking” Different

Nex AGI’s pitch centers on a framework it calls Agentic Thinking, which ties together understanding a request, planning, writing and running code, reading the result, and iterating into one continuous loop. Most models treat reasoning, tool use, and execution as separate steps. Nex-N2-Pro is trained to keep them in a single flow, which is closer to how a human engineer actually works through a problem.

Two ideas sit underneath it. Adaptive Thinking lets the model decide on its own when to reason deeply and when to act fast, which Nex AGI claims cuts thinking tokens by 30 to 50 percent on routine steps. Coherent Thinking keeps one consistent reasoning style across different task types, so the model does not behave like a different system when it switches from research to coding. Together they are meant to make long, multi-step jobs more reliable, which is exactly where most agents fall apart today.

Nex-N2.5: The September Successor

Nex AGI released the Nex-N2.5 family on September 7 and 8, 2026, three months after N2-Pro. All three sizes are open under Apache 2.0, with weights on Hugging Face and ModelScope. The Pro weights followed between September 10 and 17. Nex-N2.5-mini has about 35 billion parameters, Nex-N2.5-Pro keeps the roughly 397-billion size of N2-Pro, and Nex-N2.5-Max is a 1.6-trillion-parameter text-only model that Nex AGI calls its first complete post-training effort at trillion-parameter scale. The weights and Nex AGI’s benchmark table are on the Nex-N2.5-Pro model card.

The jump is real on paper, but every number in that table is Nex AGI’s own, run in its NexAU agent harness. Claude Opus 5 still leads most rows of the same vendor table, at 89.1 on Terminal-Bench 2.1 and 79.2 on SWE-Bench Pro, while N2.5-Max posts the best BrowseComp score in it. Artificial Analysis does not list Nex-N2.5 on its Intelligence Index yet, so test the family on your own tasks before you trust the gains.

Pricing and How to Use Nex-N2-Pro

The headline price is the easy part. Because Nex-N2-Pro is open-source under Apache 2.0, the weights are free to download from Hugging Face and ModelScope. During the launch window in June 2026, several hosted providers also offered it free, including OpenRouter on its free tier and SiliconFlow, which advertised free access for the first two weeks. Those launch offers are over. As of November 2026, OpenRouter lists Nex-N2-Pro with no active providers, SiliconFlow marks it as deprecated, and Novita’s model page is gone. Hosted access has moved to Nex-N2.5-Pro, which OpenRouter serves from Nex AGI at $0.075 per million input tokens and $0.25 per million output tokens, with Nex-N2.5-mini at $0.025 and $0.10.

You have three ways to use Nex AGI’s models today. The fastest is a hosted API, where OpenRouter now serves the N2.5 models instead of N2-Pro, so you send requests and never touch a GPU. The second is downloading the N2-Pro or N2.5 weights and running them yourself with tools like llama.cpp, Ollama, or LM Studio. The third is to skip the model-juggling entirely, which we come back to below.

The Catch: You Probably Cannot Run It Locally

Here is the reality check most launch coverage glosses over. A 397-billion-parameter model is enormous, and running Nex-N2-Pro at full BF16 precision needs roughly 794 GB of VRAM, which is data-center territory and far beyond any single consumer GPU. Even squeezed down to INT4 quantization, it still needs around 216 GB of VRAM, so a single 24 GB card like an RTX 4090 is nowhere close.

For most people that means the API route is the practical one, and the Nex-N2-mini variant is the better pick if you want something that fits on more modest hardware. If you only want strong coding answers without the infrastructure headache, Fello AI lets you put the same question to GPT, Claude, Gemini, DeepSeek, Kimi or GLM in one app on Mac, iPhone and iPad. You compare the answers with no GPUs, quantization or provider accounts to manage. Our Fello AI getting-started guide shows how the model switch works.

Nex-N2-Pro vs the Competition

So where does it land against the field? The table below sums up the trade-offs in plain terms.

If you need a free model you can host and shape to your own agent stack, Nex-N2-Pro was one of the strongest options at launch, alongside open releases like NVIDIA’s Nemotron 3 Ultra. The open field has moved fast since. On the Artificial Analysis Intelligence Index v4.3.2, read on September 25, 2026, Nex-N2-Pro scored 28.2, with at least a dozen open-weight models above it, led by MiMo-V2.6-Pro at 46.3, GLM-5.3 at 44.8 and Kimi K3 at 43.6. Artificial Analysis moved to Index v5 on October 27, 2026, which rescales every score and gives Nex-N2-Pro no number at all. Read on November 7, 2026, v5 puts Kimi K3.1 Pro first among open-weight models at 46.5, ahead of DeepSeek V4.1 Flash at 32.6 and GLM-5.3 at 31.7, both in their maximum-reasoning settings. If you just want the best answer with zero setup, a closed model or a multi-model app still wins on convenience.

Conclusion

Nex-N2-Pro is a genuine milestone for open-source AI. A free, Apache 2.0 model that matches GPT-5.5 and Claude Opus 4.7 on coding benchmarks would have sounded far-fetched a year ago, and now you can download it today. For context, the closed frontier has moved on since launch: OpenAI followed GPT-5.5 with the three-tier GPT-5.6 (Sol, Terra, and Luna) then with GPT-6 Astra and, on September 29, GPT-6.1 Sol, and Anthropic’s most capable model is now Claude Fable 5.2. The honest caveat is hardware, since running the full model takes serious infrastructure, so a hosted API for Nex-N2.5-Pro or one of the smaller mini models will be the realistic entry point for most readers.

If you build agents or write code and want an open model you control, start with Nex-N2.5-Pro on OpenRouter, which replaced N2-Pro there, and test it on your own tasks before committing to self-hosting. If you would rather skip the setup and switch between top models freely, try Fello AI’s all-in-one app instead.

FAQ

The featured image for this article was created with the help of AI.