Nex-N2.5-Max

Nex-N2.5-Max is the largest model published by Nex AGI, released with open weights on 8 September 2026 alongside the Pro and mini versions; its repository appeared a day earlier, on 7 September. The company does not pre-train: its website names DeepSeek-V4-Pro-Base as the foundation of this model, and the released configuration confirms it, declaring the DeepSeek V4 architecture, 61 layers, 384 routed experts with 6 active per token and a vocabulary of 129,280 — all of it different from the Qwen3.5 backbone used by the two smaller models. Nex AGI describes the work as its first complete post-training effort at trillion-parameter scale. The result is worth reading carefully, because it contradicts the usual assumption that the biggest model in a family is simply the best one. On the maker's own figures Max leads its siblings in coding and agent work: Terminal-Bench 2.1 86.1 against 82.7 for the Pro, SWE-Bench Pro 65.7 against 61.2, DeepSWE v1.1 65.6 against 55.8, Job Bench 53.6 against 41.4. But on OSWorld-2, the benchmark that measures actually operating a desktop computer, it scores 22.3 while the mid-sized Pro scores 56.4 and even the small mini scores 30.5. The reason is structural rather than a matter of quality: this model is text-only. The rest of the family solves such tasks by looking at the screen; Max has to work blind, and two other visual benchmarks, OSWorld-G and SWE-MM, simply have no result for it. The practical price of the model is its size. The maker's own deployment recipe calls for two nodes with sixteen H200 cards, FP8 key-value cache and a served window of 262,144 tokens — while the published configuration declares 1,048,576. That gap between the declared and the served window is the figure to watch when anyone quotes a million-token context for this model. Unlike the Pro and mini, Max was not hosted on OpenRouter at launch, so the only way to use it is to run the weights. Its Apache 2.0 licence permits that, and the licence is worth one remark: the base model it was built on comes from DeepSeek, and the derivative is being published under a permissive licence of Nex AGI's choosing. All benchmark numbers here are the maker's own, produced partly with its in-house evaluation harnesses.

▸Videos

No videos yet.

▸Reviews