Nex AGI: Nex-N2.5-Mini

Nex-N2.5-Mini is the smallest member of Nex-AGI’s Nex-N2.5 family of open-weight agentic models, released in September 2026 under the Apache 2.0 license. Built on a Qwen3.5-style multimodal Mixture-of-Experts backbone with roughly 35 billion total parameters, it extends the Nex-N2 multimodal stack with stronger computer use, web browsing, and visually grounded long-horizon workflows. The model supports adaptive reasoning modes, Qwen3-style reasoning traces, and robust function calling via the qwen3_coder tool-call parser, with a 262,144-token context window for extended agent sessions.

Nex-AGI positions Nex-N2.5-Mini for tasks where success must be verified in the environment rather than inferred from text alone: multi-step coding, GUI and browser automation, software testing from a user’s perspective, and research-style tool use. Official evaluations emphasize coding harnesses (Terminal-Bench, SWE-Bench Pro, DeepSWE), agentic productivity suites (AutomationBench, Toolathlon Verified, GDPval-AA), and browsing competence (BrowseComp), alongside multimodal computer-use benchmarks evaluated with the NexCUA harness.

Weights are published on Hugging Face and ModelScope, with hosted API access on OpenRouter. For self-hosting, Nex-AGI documents a single-node deployment using two H100 GPUs with tensor parallelism (tp=2) and the customized nexagi/sglang Docker image, including the repository chat template and recommended sampling settings (temperature 0.7, top_p 0.95, top_k 40).