Nex-AGI Ships 35B Nex-N2.5-mini Agent Model Under Apache 2.0

Nex-AGI has posted Nex-N2.5-mini on Hugging Face, releasing the 35B BF16 checkpoint under Apache 2.0 with deployment recipes for vLLM, SGLang and a bundled Docker image. The card presents it as the entry point in a three-tier lineup that scales up to Nex-N2.5-Max, described on the page as "a 1.6 trillion parameter Mixture-of-Experts (MoE) text-only model."

On the reported agent leaderboard, the mini card lists 73.4 on Terminal-Bench 2.1, 43.8 on SWE-Bench Pro, 54.6 on Toolathlon Verified and 83.4 on BrowseComp for text-agent tasks, with 71.2 on OSWorld-Verified and 63.4 on WebArena-Verified for computer-use. Each row is paired with a "Top Performer" column citing labels like "Claude Opus 5" at 89.1 Terminal-Bench, "Kimi-K3" at 76.5 Toolathlon and "Qwen3.8-Max" at 86.1 OSWorld, model names that do not correspond to any publicly shipped release, and no third-party reproduction of the numbers has surfaced yet.

Under the hood, the card describes the model as a "Qwen3.5 MoE base with multimodal extensions" wired for function calling through a qwen3_coder parser and an adaptive reasoning_effort control with none, medium and high modes. Recommended sampling is temperature 0.7, top-p 0.95, top-k 40. Weights are mirrored on ModelScope and hosted access runs through OpenRouter.

Open-agent drops have been arriving in near-daily cadence, with 414 agent stories in our tracker over the last 90 days; this one lands alongside our same-day coverage of Tencent's Gander omni-interaction agent.