--- title: 'Nex-N2-Pro: Open-Weight Coder vs GPT-5.5 (2026)' description: Nex-N2-Pro is a free, open-weight 397B MoE coding model from Nex AGI, built on Qwen3.5. We check its benchmarks against GPT-5.5, Opus, and rival open models. url: https://nerdleveltech.com/nex-n2-pro-open-weight-coding-benchmarks language: en date: '2026-06-10' category: ai-ml tags: - nex-n2-pro - nex agi - open-weight models - ai coding models - llm benchmarks - qwen3.5 - agentic ai markdown_url: https://nerdleveltech.com/nex-n2-pro-open-weight-coding-benchmarks.md attribution: 'Source: NerdLevelTech — https://nerdleveltech.com/nex-n2-pro-open-weight-coding-benchmarks' terms: https://nerdleveltech.com/llms.txt --- # Nex-N2-Pro: Open-Weight Coder vs GPT-5.5 (2026) Nex-N2-Pro is a free, open-weight 397-billion-parameter coding model from Nex AGI, post-trained on Alibaba's Qwen3.5.[^1][^2] On its own benchmarks it leads the open-weight pack and edges GPT-5.5 on SWE-Bench Pro, but it trails the closed frontier on the hardest agentic and reasoning tasks.[^1][^7] Released June 2, 2026 and now spreading across free endpoints, it is the latest entry in an early-June surge of open-weight coding models — arriving a day after [MiniMax M3](minimax-m3-open-weight-coding-sparse-attention).[^3][^10] Two things set it apart: its headline feature is an "Agentic Thinking" framework that decides on its own how hard to reason, and — unlike M3, whose weights were still pending at launch — you can actually download Nex-N2-Pro right now.[^1][^2] ## TL;DR Nex-N2-Pro is the open-weight, agentic coding model from **Nex AGI**, an open-source alliance initiated by the Shanghai Innovation Institute.[^1][^5] It is a Mixture-of-Experts model with **397 billion total parameters but only 17 billion active per token**, post-trained on Alibaba's `Qwen3.5-397B-A17B` base, with a 262K-token context window and image input.[^2][^3][^4] Its signature trick is **Adaptive Thinking**, which auto-adjusts reasoning depth per step and, Nex says, cuts thinking tokens by 30–50% versus always-on reasoning at equal or better quality.[^3] On Nex's own benchmark table, Nex-N2-Pro scores **80.8% on SWE-Bench Verified, 58.8% on SWE-Bench Pro, and 75.3 on Terminal-Bench 2.1**, narrowly beating GPT-5.5 on SWE-Bench Pro and Claude Opus 4.7 on Terminal-Bench, while leading or matching rival open models like MiniMax M3, DeepSeek-V4-Pro, GLM-5.1, and Kimi-K2.6 on most rows.[^1] Three caveats matter: every number is **self-reported on Nex's own harness**; the comparison baseline is **Opus 4.7, not the newer Opus 4.8**; and on the toughest agentic tests the gap to the closed frontier is wide (DeepSWE: 33.6 vs GPT-5.5's 70).[^1][^7][^8] The weights are free under Apache 2.0, and it runs free on OpenRouter and SiliconFlow during launch — but self-hosting the full model needs roughly two nodes of 8×H100.[^1][^6] ## What is Nex-N2-Pro? Nex-N2-Pro is a large language model built for coding and agentic work, released and open-sourced on June 2, 2026 by **Nex AGI** (also styled "Nex").[^1][^3] Nex is not a conventional startup but an open-source alliance "initiated by the Shanghai Innovation Institute," developing in the open under the `nex-agi` organization on GitHub and Hugging Face, with partners including Shanghai Qiji Zhifeng, Mosi Intelligence, and KuafuAI.[^5] The group ships a full agent stack — models, an agent framework (NexAU), data pipelines, and training infrastructure — of which Nex-N2 is the flagship model.[^5] Architecturally, Nex-N2-Pro is a sparse **Mixture-of-Experts (MoE)** model with **397 billion total parameters and 17 billion active per token**, so its inference cost behaves closer to a 17B dense model than its headline size suggests.[^2][^3] It accepts text and images as input and produces text, supports a **262,000-token context window**, and ships under the permissive **Apache 2.0** license with downloadable weights on Hugging Face and ModelScope.[^2][^3] The crucial detail — and one the marketing soft-pedals — is that Nex-N2-Pro is **post-trained on Alibaba's `Qwen3.5-397B-A17B`**, not a base model Nex built from scratch.[^1][^2] The 397B/17B architecture, the native multimodality, and the long-context efficiency all come from Qwen3.5, Alibaba's flagship open-weight MoE from February 2026, which fuses linear attention (via Gated Delta Networks) with a sparse MoE.[^4] Nex AGI's actual contribution sits on top: the agentic post-training. The previous generation, Nex-N1, was built the same way on a different base — DeepSeek-V3.1 — so swapping in Qwen3.5 for N2 is the lineage here.[^3] ## What is Agentic Thinking (and Adaptive Thinking)? **Agentic Thinking** is Nex's framework for unifying reasoning, tool use, and environment execution into a single loop, instead of bolting them together as separate capabilities.[^1] It is the headline idea behind the model, and it has two halves.[^1] **Adaptive Thinking** lets the model decide, on its own, when to think and how deeply — running simple actions fast while reserving thorough reasoning for the decisions that matter. Nex's specific claim is that this cuts thinking tokens by **30–50% on routine steps** versus always-on reasoning, with equal or better task performance.[^3] That is the part worth caring about: long agentic runs burn most of their cost on reasoning tokens, so a model that reasons only when it needs to is cheaper to operate at the same quality — if the claim holds up under independent testing. **Coherent Thinking** is the second half: carrying one consistent reasoning style across general tasks and diverse agentic work, so capability transfers cleanly between, say, terminal execution and multimodal generation.[^1] In practice both behaviors are exposed through standard tooling — Nex-series models emit explicit reasoning traces and support function calling, using the same `qwen3` reasoning parser and `qwen3_coder` tool-call parser as their Qwen base, and they advertise plug-and-play use with Claude Code, Cursor, and other agent harnesses.[^2][^3] ## Nex-N2-Pro benchmarks: strong, but self-reported On Nex's published evaluation suite, Nex-N2-Pro's coding numbers are genuinely good. It posts **80.8% on SWE-Bench Verified** (fixing real bugs in real repositories), **58.8% on SWE-Bench Pro**, **75.3 on Terminal-Bench 2.1**, and **1585 on GDPval** — OpenAI's benchmark for real-world, economically valuable knowledge work scored by expert graders across 44 occupations.[^1][^9] Nex frames the model as keeping "pace with top-tier models such as GPT-5.5 and Opus 4.7."[^1] Here is the context most coverage drops. **Every one of those numbers comes from Nex AGI's own benchmark table** — its harness, its scaffolding, its choice of baselines.[^1] That is standard practice for a model launch, but it makes the results claims awaiting independent replication, not settled facts. The table even shows the seams: it lists GPT-5.5 at 82.9% on SWE-Bench Verified, whereas OpenAI officially reports GPT-5.5 at **88.7%** on the same benchmark, and independent harnesses like Vals.ai put GPT-5.5 lower still at ~82.6%.[^7][^11] In other words, the comparison column is a mix of harness conditions, which is exactly why a self-run table should be read as a strong opening bid rather than a verdict. | Benchmark (Nex-reported) | Nex-N2-Pro | GPT-5.5 | Opus 4.7 | MiniMax M3 | DeepSeek-V4-Pro | |---|---|---|---|---|---| | SWE-Bench Verified | 80.8 | 82.9 | 87.6 | 80.5 | 80.6 | | SWE-Bench Pro | 58.8 | 58.6 | 64.3 | 59.0 | 55.4 | | Terminal-Bench 2.1 | 75.3 | 83.4 | 69.7 | 66.0 | 72.0 | | DeepSWE | 33.6 | 70 | 54 | — | 8 | | GDPval | 1585 | 1769 | 1753 | — | 1554 | | GPQA Diamond | 90.7 | 93.6 | 94.2 | — | 90.1 | All figures as published by Nex AGI.[^1] Higher is better in every row; GDPval is a score, the rest are percentages. ## Nex-N2-Pro vs GPT-5.5 and Claude Opus: the honest read Strip away the "keeps pace" framing and the picture is more specific. Against **GPT-5.5**, Nex-N2-Pro edges ahead on SWE-Bench Pro (58.8 vs 58.6) but trails on SWE-Bench Verified (80.8 vs 82.9), on Terminal-Bench 2.1 (75.3 vs 83.4), and badly on DeepSWE (33.6 vs 70).[^1] Against **Claude Opus 4.7**, it wins Terminal-Bench 2.1 (75.3 vs 69.7) but loses SWE-Bench Verified by nearly seven points (80.8 vs 87.6) and DeepSWE by twenty (33.6 vs 54).[^1] The selective wins are real; so are the losses. Two structural caveats deepen the point. First, the comparison ceiling is **Claude Opus 4.7**, but Anthropic shipped **Claude Opus 4.8** on May 28, 2026 — before Nex-N2-Pro's June 2 launch — and Opus 4.8 reports roughly **88.6% on SWE-Bench Verified**.[^8] Measured against the Anthropic model that was actually current, the gap is wider than the table implies. It is the same flattering-baseline pattern we just flagged in the [MiniMax M3 launch](minimax-m3-open-weight-coding-sparse-attention), where M3 was also benchmarked against Opus 4.7 rather than 4.8.[^10] Second, the **DeepSWE** and **GDPval** rows are where the closed frontier pulls clearly ahead: a 33.6 on DeepSWE against GPT-5.5's 70 is not "near-parity" on hard agentic coding, it is roughly half. (We dug into how brittle and gameable that benchmark is in [our DeepSWE breakdown](deepswe-benchmark-claude-cheating-gpt-5-5).) ## The open-weight pack: where Nex-N2-Pro actually leads The fairer frame is not Nex-N2-Pro versus the closed frontier — it is Nex-N2-Pro versus the other open-weight challengers, and there it looks like a leader. On Nex's table it matches or beats **MiniMax M3, DeepSeek-V4-Pro, GLM-5.1, and Kimi-K2.6** on most rows: 80.8 on SWE-Bench Verified sits at the top of a tight open-weight cluster (M3 80.5, DeepSeek-V4-Pro 80.6, Kimi-K2.6 80.2); on Terminal-Bench 2.1 its 75.3 clears MiniMax M3's 66.0 and DeepSeek-V4-Pro's 72.0; and on DeepSWE its 33.6, while far behind the closed models, towers over the open field (Kimi-K2.6 24, GLM-5.1 18, DeepSeek-V4-Pro 8).[^1] That positioning is the story. Over the past few months we have watched [Chinese open-weight coding labs run an aggressive cost-and-capability war](china-open-weight-coding-models-cost-war), from [DeepSeek V4](deepseek-v4-open-source-frontier-million-token-context) to [GLM-5.1](glm-5-1-open-source-beats-gpt-coding-benchmarks) to [Kimi K2.6](kimi-k2-6-300-agent-swarm-open-weight-frontier-coding). Nex-N2-Pro's pitch is to take one of the strongest open bases — Alibaba's Qwen3.5 — and squeeze more agentic performance out of it through post-training, then give it away. If the numbers replicate, it is the new front-runner of the open-weight coding pack, even though it is not a genuine threat to GPT-5.5 or Opus 4.8 on the hardest tasks. ## Nex-N2-Pro pricing: free now, but heavy to host The price is the easy part: Nex-N2-Pro is **free**. The weights are Apache 2.0 and downloadable from Hugging Face and ModelScope, and during the launch window the model runs at no cost on hosted endpoints — **$0 input and $0 output on OpenRouter's free tier** (rate-limited to roughly 50 requests/day and 20/minute) and free early access on SiliconFlow.[^2][^6] For a model posting frontier-adjacent coding scores, free hosted access is a strong way to drive adoption. The catch is hardware. "Open weight" does not mean "runs on your laptop." Nex's own deployment guide launches Nex-N2-Pro across **two nodes of 8× H100 GPUs** (tensor-parallel 16) using its customized SGLang fork.[^1] That is a serious cluster — far beyond what an individual can self-host — so for most people "free" in practice means the rate-limited hosted tiers, not local inference. The smaller sibling, **Nex-N2-mini** (built on `Qwen3.5-35B-A3B-Base`), is the more self-hostable option Nex points to, running on a single 2× H100 box, though it gives up real ground on the benchmarks (74.4 vs 80.8 on SWE-Bench Verified).[^1] ## FAQ **What is Nex-N2-Pro?** Nex-N2-Pro is an open-weight, agentic coding model released June 2, 2026 by Nex AGI, an open-source alliance initiated by the Shanghai Innovation Institute. It is a 397B-parameter Mixture-of-Experts model (17B active per token) post-trained on Alibaba's Qwen3.5, with a 262K-token context window and image input.[^1][^2][^3] **Is Nex-N2-Pro free?** Yes. The weights are released under Apache 2.0 and free to download from Hugging Face and ModelScope, and during launch the model is free to run on OpenRouter's free tier ($0/$0, rate-limited) and on SiliconFlow's early access.[^2][^6] **Is Nex-N2-Pro better than GPT-5.5?** On Nex's own benchmarks it narrowly beats GPT-5.5 on SWE-Bench Pro (58.8 vs 58.6) but trails on SWE-Bench Verified, Terminal-Bench 2.1, and especially DeepSWE (33.6 vs 70). The numbers are self-reported and not independently verified, so treat "better" as benchmark-specific and provisional.[^1][^7] **What model is Nex-N2-Pro based on?** It is post-trained on Alibaba's `Qwen3.5-397B-A17B`, the flagship open-weight Qwen MoE from February 2026. The architecture, multimodality, and efficiency come from Qwen3.5; Nex AGI's contribution is the agentic post-training (Agentic Thinking).[^2][^4] **Can you download Nex-N2-Pro?** Yes — unlike MiniMax M3, whose weights were still pending at its launch, Nex-N2-Pro's weights are live on Hugging Face and ModelScope. But the full model needs roughly two 8×H100 nodes to self-host, so most users will rely on hosted endpoints.[^1][^2] **How many parameters does Nex-N2-Pro have?** 397 billion total parameters, with 17 billion active per token thanks to its Mixture-of-Experts design — the same 397B/17B configuration as its Qwen3.5-397B-A17B base.[^2][^4] ## Bottom line Nex-N2-Pro is a real step for open-weight coding, wrapped in a launch that oversells the comparison. The Agentic Thinking idea — reasoning only as deeply as a step requires — is a sensible answer to the cost of long agent runs, and giving the weights away free under Apache 2.0 is the kind of move that drives fast adoption.[^1][^3] But the benchmark narrative leans on self-run tests against Opus 4.7, a baseline Anthropic had already replaced with Opus 4.8, and the hardest agentic tests (DeepSWE, GDPval) show the closed frontier is still comfortably ahead.[^1][^8] The honest summary: Nex-N2-Pro is probably the strongest open-weight coding model you can download today, and a weaker claimant to outright GPT-5.5 parity than its own charts suggest. Independent benchmarks — and the experience of actually running it inside a real agent harness — will settle which half of that sentence matters more. --- *Related reading: [MiniMax M3: open-weight coding at 1/10 the cost](minimax-m3-open-weight-coding-sparse-attention), [DeepSeek V4: open-weight frontier at 1/7 the cost](deepseek-v4-open-source-frontier-million-token-context), and [China's open-weight coding wave](china-open-weight-coding-models-cost-war).* ## Related reads - [MiniMax M3: Open-Weight Coding at 1/10 the Cost (2026)](https://nerdleveltech.com/minimax-m3-open-weight-coding-sparse-attention) [^1]: Nex AGI, "Nex-N2: An agentic model with Agentic Thinking" — official GitHub README (release, Agentic Thinking / Adaptive Thinking framework, full benchmark table, Qwen3.5 base, two-node 8×H100 deployment guide). https://github.com/nex-agi/Nex-N2 [^2]: "nex-agi/Nex-N2-Pro," Hugging Face model card (Apache-2.0 license, 397B params, MoE `qwen3_5_moe` architecture, image-text-to-text, benchmark table, parsers). https://huggingface.co/nex-agi/Nex-N2-Pro [^3]: "Nex-N2-Pro — Model Info, Parameters, Benchmarks," SiliconFlow (created Jun 2, 2026; 262K context window; FP8; Apache-2.0; Adaptive Thinking 30–50% thinking-token claim; prior-generation DeepSeek-V3.1-Nex-N1). https://www.siliconflow.com/models/nex-n2-pro [^4]: "Alibaba's latest flagship Qwen3.5 models are open-weights MoE performers," DeepLearning.AI The Batch, and "Qwen3.5-397B-A17B — Everything you need to know," Artificial Analysis (Feb 2026 release, 397B total / 17B active, hybrid linear-attention + sparse MoE, native multimodal, Apache 2.0). https://www.deeplearning.ai/the-batch/alibabas-latest-flagship-models-are-open-weights-moe-performers-in-sizes-from-less-than-1b-parameters/ [^5]: "Nex: Nexus of Agentic Intelligence," Nex AGI (alliance initiated by the Shanghai Innovation Institute; partners and full-stack agent ecosystem — models, NexAU framework, data and RL infrastructure). https://nex-agi.com/en/ [^6]: "Nex AGI: Nex-N2-Pro (free) — API Pricing & Providers," OpenRouter ($0 input / $0 output free tier; free-tier rate limits). https://openrouter.ai/nex-agi/nex-n2-pro:free/pricing [^7]: "SWE-Bench Leaderboard," marc0.dev (GPT-5.5 #1 at 88.7% on SWE-Bench Verified, OpenAI-reported, released Apr 23, 2026; Claude Opus 4.7 87.6%, Apr 16, 2026), and Vals.ai independent harness (GPT-5.5 ~82.6%, Opus 4.7 ~82.0%) showing harness variance. https://www.marc0.dev/en/leaderboard [^8]: "Anthropic's Claude Opus 4.8 is here," VentureBeat (Opus 4.8 released May 28, 2026; ~88.6% on SWE-Bench Verified). https://venturebeat.com/technology/anthropics-claude-opus-4-8-is-here-with-3x-cheaper-fast-mode-and-near-mythos-level-alignment [^9]: "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks," OpenAI (benchmark across 44 occupations in the top 9 GDP sectors, expert blind grading). https://openai.com/index/gdpval/ [^10]: "New Models Today — AI & LLM Releases," Price Per Token (Nex-N2-Pro free listing on OpenRouter; MiniMax M3 release context). https://pricepertoken.com/news/model-releases [^11]: "SWE-bench Verified," Vals.ai independent benchmark harness (GPT-5.5 82.60%, Claude Opus 4.7 82.00%). https://www.vals.ai/benchmarks/swebench