Nex-N2-mini

Nex-N2-mini is the smaller of the two models Nex AGI open-sourced in June 2026 as its second generation, announced in the same model card as the larger Nex-N2-Pro. Both are post-trained on the Qwen3.5 series rather than built from scratch: the mini on Qwen3.5-35B-A3B-Base, the Pro on Qwen3.5-397B-A17B. The weights carry 35.1 billion parameters in bfloat16, spread over 40 layers and 256 experts of which 8 fire per token, and Nex AGI's contribution is the post-training, not the architecture. The pitch is the same Agentic Thinking framework as the Pro edition — Adaptive Thinking, which lets the model decide when to reason and how deeply, and Coherent Thinking, meant to keep one reasoning paradigm across ordinary and agentic work. What the mini adds is reach: where the Pro edition needs two nodes of eight H100 accelerators, the mini is documented as running on a single machine with two of them, which is the difference between a data-centre deployment and a workstation. The maker's own comparison table is honest about the cost of that. Against the Pro edition the mini holds close on some agentic rows — BrowseComp 74.1 against 83.7, TAU3 65.9 against 71.1, SWE-Bench Verified 74.4 against 80.8 — and collapses on the hardest ones. On DeepSWE it scores 8.0 where the Pro edition reaches 33.6 and GPT-5.5 reaches 70; on Apex it scores 9.4 against the Pro's 36.5. Terminal-Bench 2.1 lands at 60.7, SWE-Bench Pro at 50.2, GPQA Diamond at 82.6, GDPval at 1402. It is a model that stays usable on cheap hardware for routine agent work and stops being competitive the moment a task gets genuinely long. Two things in the released files are absent from the model card. The configuration sets a 262,144-token context window, a figure Nex AGI never states in writing. And the model is not text-only: the weight index holds 333 vision tensors out of 1,026, the repository ships a Qwen3-VL image processor, and the configuration defines image and video tokens — Hugging Face accordingly files the model as image-text-to-text. Nex AGI documents none of this, claims nothing about images, and publishes a benchmark table that is entirely textual, so the vision path is best read as inherited from the Qwen3.5 base and left undeclared rather than as an advertised capability. Weights went out on Hugging Face and ModelScope under Apache 2.0, with no first-party price list.

▸Videos

No videos yet.

▸Reviews