Nex-N2-Pro: How the Open-Model Frontier Shrunk 75 % in Six Weeks

Six weeks, two orders of magnitude

On April 22, 2026, DeepSeek launched DeepSeek-V4-Pro – the largest open-weight model ever released: 1.6 trillion parameters, 49B active, MoE, 1M-token context. A technical statement – and for many, proof that the open frontier had finally caught up with the closed one.

On June 5, 2026, nex-agi released Nex-N2-Pro: 397 billion parameters, 17B active, Apache 2.0, post-trained on Qwen3.5-397B-A17B. One quarter the total parameters. About one third the active ones. And on the benchmarks that matter for agentic work, on par with or ahead of V4-Pro.

The benchmarks that actually matter

Nex-N2 is explicitly an agentic model – not "just another chat." The official numbers on the benchmarks that stand for tool use, coding and long-horizon execution:

This is not "pretty good for an open model." This is frontier level at one quarter the size – with a license no one will drag you to court over.

What explains this curve technically

Three levers are running in parallel, reinforcing each other:

1. Better MoE architectures. 17B active instead of 49B means: less VRAM per token, higher throughput, lower latency. The choice of which expert fires is significantly smarter today than it was in 2024.
2. Post-training on strong bases. Nex-N2-Pro is post-trained on Qwen3.5-397B-A17B. The community no longer builds from scratch; it stacks specialization on top of proven pretrained cores.
3. Agentic thinking as a first-class capability. Instead of reactively attaching chain-of-thought, reasoning is part of the architecture and training. The model decides when and how deeply to think – similar to Mistral Medium 3.5 and GPT-5.4.

Translated: We do not get slightly more performance per parameter each year – we get it in leaps.

What used to need a cluster, runs locally tomorrow

A very rough but honest curve of the "largest open frontier models":

2024 Q4   Llama-3.1-405B Dense           405 B
2025 Q2   DeepSeek-V3                    671 B  (37B active)
2025 Q4   Kimi K2 / Qwen3.5             ~1,000 B
2026 Q2   DeepSeek-V4-Pro                1,600 B (49B active)
2026 Q2   Nex-N2-Pro                       397 B (17B active)

The point is not "small beats large." The point is: the efficient frontier shifts downward while the maximum frontier keeps moving upward. The tension between "what can maximally be done" and "what do you actually need" keeps growing – in favor of builders.

Concretely: What ran on a multi-node H100 cluster a year ago now runs on a single 8×H100 node with the right quantizations – and with aggressive GGUF/AWQ increasingly on workstations or edge boxes. Frontier level is migrating from the hyperscaler data center to the desktop.

Who wins – and who does not?

The winners

• Builders with sovereignty requirements. Healthcare, public sector, banking, pharma, industry. Those who cannot push data through US APIs get a model that delivers instead of merely "suffices."
• EU companies under GDPR pressure. Self-hosted, Apache 2.0, clear supply chain. No more "we have the DPA, trust us" layer needed.
• AI-first builders. Anyone running agents in production can switch the default model without rebuilding the architecture stack. Tool-calling quality becomes a swappable component.
• Edge and on-device scenarios. 17B active is the threshold where "runs on a workstation" becomes realistic – with genuine reasoning depth, not just as a toy.

The losers (at least at the margin)

• Hyperscaler pricing for pure inference. When a 397B model with an Apache 2.0 license can do frontier coding, willingness to pay $ for GPT-class inference drops – except for the exact workloads where closed models still truly lead.
• Closed "open-but-not-really" licenses. Modified MIT with MAU thresholds or acceptable-use clauses suddenly look expensive when real Apache 2.0 sits right next to them.
• Pure model-license business models. Value creation is finally shifting to the tool layer – Cursor, Vibe, Claude Code, IDE integrations, agentic runtime, eval stack. Anyone just selling weights is selling a commodity.

What we are doing concretely

At Till Freitag, we are rotating Nex-N2-Pro into our internal eval suite this week – alongside Medium 3.5, Claude Sonnet 4.5 and Kimi K2.5. Three questions we want answered:

1. Tool-calling accuracy on real monday.com and RevOps workflows.
2. Long-horizon stability over 50+ steps – Terminal-Bench reads well, but the real question is drift in multi-step agents.
3. Inference economics self-hosted on 8×H100 NVL vs. managed via SiliconFlow – at what point does self-operation pay off?

First impressions will land in a follow-up within the next two weeks. Want to compare along? Model card is on Hugging Face, playground at SiliconFlow.

What to do now

1. If you are running DeepSeek-V4-Pro or Llama 4 Maverick today: Evaluate Nex-N2-Pro on your own tasks. Not on marketing benchmarks – on your workflows.
2. If you do not have an open-weights strategy yet: Get in now. The curve of the next twelve months will renegotiate the pricing of your closed-source contracts.
3. If you are building in regulated industries: Self-hosting is shifting from premium path to default path. Plan your infra accordingly.
4. If you only need chat: Stay with your managed provider. This wave is built for agents, coding and long-horizon tasks.

Bottom line

Six weeks ago, "largest open model" meant "1.6 trillion parameters." Today, the largest open model leading the agentic benchmarks is one quarter that size. This half-life is not an anomaly – it is the new pace.

The honest answer to "who wins" is: everyone who wants to own their AI. Frontier level is no longer exclusive hyperscaler territory. It is a release you can download this week, operate in license compliance, and embed in your product tomorrow.

The exciting question is no longer whether open models win. It is what you build when frontier AI sits on your desktop.

---

→ Mistral 3, Large 3 & Vibe: Europe's AI hope back in the game → Open-source LLM comparison: 20+ models at a glance → GGUF, GGML, Safetensors – formats for local AI → Our AI services at Till Freitag