321.3B total / approximately 18B active (sparse MoE; FP8 checkpoint)
1,048,576 tokens in the released configuration; 262,144 tokens through OpenRouter
Apache 2.0
Open-weight multimodal agentic model post-trained from GLM-5.3-Flash for computer use, coding, long-horizon workflows, and tool-driven tasks. Nex-N2.6 retains the GLM-5.3-Flash architecture with approximately 321.3 billion total parameters and an approximately 18-billion-parameter active path. Its 45-layer, 4,096-dimensional language backbone contains three initial dense feed-forward layers followed by 42 sparse Mixture-of-Experts layers. Each sparse layer contains 288 routed experts with eight selected per token plus one shared expert. The attention stack alternates groups of three linear-attention layers with one DeepSeek Sparse Attention layer, producing 34 linear and 11 sparse-attention layers in total. Other architectural features include four-branch Manifold-Constrained Hyper-Connections, IndexPool compression for sparse-attention retrieval, and one Multi-Token Prediction layer. A 24-layer, 1,024-dimensional vision encoder with spatial and temporal patching supports image and video inputs, while generated output remains text. The released configuration specifies a 1,048,576-token context window, although the hosted OpenRouter endpoint exposes 262,144 tokens. The bundled chat template supports low, high, and max reasoning-effort settings, defaults to max, and supports structured tool use. Nex AGI says this release focuses on eliminating repetitive reasoning and improving adherence to the tools actually supplied by an agent environment. Vendor-reported results include 83.0 on OSWorld, 84.1 on Terminal-Bench 2.1, and 45.7 on AutomationBench. In Nex AGI's internal replay evaluation, reasoning repetition fell from 16.0% for Nex-N2.5-Pro to 1.6%, the overall non-repetition rate reached 97.8%, and valid tool selection reached 100% both overall and among responses that made tool calls. The downloadable checkpoint uses blockwise E4M3 FP8 weights while retaining selected modules at BF16 or FP32 precision. It contains approximately 314.4B FP8, 6.92B BF16, and 3.6M FP32 parameters across 129 safetensor shards totaling approximately 328.34 GB. Nex AGI recommends serving it through an FP8-compatible SGLang configuration with GLM-5.3-Flash support. The checkpoint is released under Apache 2.0.