1.6T total (Max); 35B total (mini)
1,048,576 tokens configured for Max (official serving recipe uses 262,144); 262,144 tokens for mini
Apache-2.0
Official open-weight portion of the Nex-N2.5 agent-model family, covering the downloadable Nex-N2.5-Max and Nex-N2.5-mini checkpoints. The same-day Nex-N2.5-Pro variant is announced and available through hosted inference, but is excluded from this release grouping because its official repositories currently contain only a model card and figures rather than checkpoint shards or a configuration. Nex-N2.5-Max is a text-only 1.6-trillion-parameter Mixture-of-Experts model based on a DeepSeek-V4-style architecture and represents Nex AGI's first complete trillion-scale post-training effort. Its configuration specifies 61 layers, a 7,168-dimensional hidden state, 384 routed experts plus one shared expert, six selected routed experts per token, three initial hash-MoE layers, Hyper Connections with multiplicity four, and alternating heavily compressed and compressed sparse attention. The checkpoint configures 1,048,576 positions using YaRN factor 16 from an original 65,536-token window, although Nex AGI's published multi-node serving recipe currently caps runtime context at 262,144. Expert weights use blockwise FP8 while other components remain at higher precision; the repository totals approximately 1.65 TB across 644 safetensors shards. Nex-N2.5-mini is a 35B multimodal Qwen3.5-style sparse MoE with 40 language layers, three linear-attention layers for every full-attention layer, 256 routed experts with eight selected per token, a shared expert, and a native 262,144-token context. Its visual stack uses a 27-layer, 1,152-dimensional ViT with spatial and temporal patching, enabling image-grounded computer use and agent workflows. The mini checkpoint totals approximately 70.2 GB across 16 BF16 safetensors shards. Both released variants support non-thinking, adaptive-thinking, and always-thinking modes through reasoning_effort, along with structured function calling for agent harnesses. Nex AGI's vendor-reported results include 86.1 on Terminal-Bench 2.1, 65.7 on SWE-Bench Pro, 65.6 on DeepSWE v1.1, and 92.6 on BrowseComp for Max; mini reports 73.4, 43.8, 36.1, and 83.4 respectively, plus 71.2 on OSWorld-Verified and 52.9 on Vision2Web. These scores use Nex AGI's stated NexAU and NexCUA harnesses and should be interpreted as vendor evaluations rather than controlled independent comparisons.