Nex-N2.6: A 321B Open-Weight Agentic Model That Claims 83.0 on OSWorld

A humanoid android robot

Photo: Hippopx (public domain). Illustrative image.

What happened. Nex-AGI has published Nex-N2.6 on Hugging Face — a roughly 321-billion-parameter multimodal model built for agentic work, released on 10 November 2026 with its repository updated on 14 November. It is built on GLM-5.3-Flash, ships as FP8 weights, and carries an Apache 2.0 licence, which makes it one of the largest openly licensed agentic models available to run and build on.

The numbers Nex-AGI is leading with. The model card claims 83.0 on OSWorld, the benchmark for computer-use agents, against 82.2 for its predecessor Nex-N2.5-Pro and 78.0 for the base GLM-5.3-Flash. On Terminal-Bench 2.1 it claims 84.1. Those are Nex-AGI’s own figures, measured on its own terms, and no independent test of the FP8 export exists yet — so treat them as the vendor’s scorecard, not a settled ranking. The more interesting claim is about reliability rather than capability: reasoning repetition falls from 16.0% to 1.6% versus Nex-N2.5-Pro, and the model hits 100% tool-list adherence in the company’s internal evaluation. For an agent that runs long, multi-step tasks, a model that stops repeating itself and actually calls the tools it was given matters more than a point on a benchmark.

The practical details. Nex-N2.6 has a 262,000-token context window and is listed on OpenRouter at about $0.12 per million input tokens and $0.40 per million output tokens, which puts it in the cheap tier for a model of this size. It is aimed at computer use, coding and long-horizon tool tasks — the same territory where Thinking Machines’ Inkling 2 Small landed this week and where Edge0-145B claims a 145B model can run on a single GPU. The open-weight agentic lane is getting crowded fast: the question for builders is no longer whether a capable open model exists, but which one’s licence, size and reliability trade-offs fit the job — and Kimi K3.1-Pro is the other open-weight release to compare it against.

Sources

  • Nex-AGI
  • Nex-N2.6
  • open weights
  • AI agents
  • OSWorld

Related articles