Nex-AGI's New Agent Model Runs on Qwen's Architecture

Nex-AGI released three open-weight agent models under an Apache 2.0 license on Sept. 8, 2026, including a 1.6-trillion-parameter mixture-of-experts model the company calls its first complete post-training effort at that scale.

The company published weights for Nex-N2.5-mini, Nex-N2.5-Pro and Nex-N2.5-Max to Hugging Face and ModelScope the same day, according to the models' own Hugging Face cards and the project's GitHub repository.

What the smallest model is actually built on

Nex-N2.5-mini's configuration file, published alongside its weights, lists its architecture as Qwen3_5MoeForConditionalGeneration and its model type as qwen3_5_moe: the same class Alibaba uses for its own Qwen3.5 mixture-of-experts models. The file shows 256 routed experts with eight active per token, a hidden size of 2,048 and 40 transformer layers. Nex-AGI's own announcement material describes the mini model only as part of "a next-generation family of agentic models," without naming the base architecture it was built on.

Model	Parameters	Context window	License
Nex-N2.5-mini	35B (~3B active)	262,144 tokens	Apache 2.0
Nex-N2.5-Pro	397B	not disclosed	Apache 2.0
Nex-N2.5-Max	1.6T (MoE)	262,144 tokens	Apache 2.0

Where the benchmarks actually land

Nex-N2.5-Pro scores 82.7 on Terminal-Bench 2.1, an agentic coding-and-operations benchmark, and 61.2 on SWE-Bench Pro, against 73.4 and 43.8 for the smaller mini model, according to Nex-AGI's published benchmark tables. On OSWorld-Verified, which grades a model's ability to control a real desktop environment, Pro scores 82.2 versus mini's 71.2.

1. Terminal-Bench 2.1 (mini)73.4 score
2. Terminal-Bench 2.1 (Pro)82.7 score
3. SWE-Bench Pro (mini)43.8 score
4. SWE-Bench Pro (Pro)61.2 score

Source: Nex-N2.5-mini and Nex-N2.5-Pro Hugging Face model cards, accessed 2026-09-12

Pro is the only one of the three models whose context window Nex-AGI has not published; mini and Max both list 262,144 tokens, roughly 200,000 words, on their Hugging Face cards. The company markets all three specifically for controlling a computer or browser rather than answering chat questions, a use case it groups under "visually grounded agentic capabilities."

What it takes to run the smallest one

Nex-N2.5-mini needs a single two-H100 GPU node as its reference configuration, according to a deployment writeup by OrcaRouter, an inference-hosting site that reviewed the release the same week. Hosted, API-style access to the mini and Pro tiers is also available through OpenRouter, Nex-AGI said, without Nex-AGI having to publish the raw weights for every user who wants to try them.

Nex-N2.5 arrives days after Qwen's own 27-billion-parameter Qwen3.8 model became one of Hugging Face's most-downloaded releases, and in the same month that DeepSeek shipped its own 552-billion-parameter mixture-of-experts model, V4.1-Flash. Nex-AGI has not said whether it trained Nex-N2.5-mini on top of a published Qwen3.5 checkpoint or reproduced the architecture independently, and neither the model card nor the GitHub repository discloses the training data behind any of the three releases.