Nex-AGI launches 1.6T N2.5 agent family, leaves Pro weights pending

Nex-AGI, a collaboration involving the Shanghai Innovation Institute, Shanghai Qiji Zhifeng, Mosi Intelligence and Kuafu Technology, released the N2.5 agent family on September 8, 2026, comprising the 35B Mini, 397B Pro and 1.6T Max models, with Pro weights still marked as coming soon. The models target computer use, browsing, reasoning and agent tasks, with Mini and Pro supporting multimodal input and Max being text-only, and deployment requirements range from two H100s for Mini to 16 H200s across two nodes for Max.

Nex-AGI launches 1.6T N2.5 agent family, leaves Pro weights pending Mini and Pro target multimodal computer use, while text-only Max requires far more accelerator capacity and Pro remains unavailable for download. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Nex Ecosystem https://x.com/NexEcosystem/status/2097341149405151287 Why it matters N2.5 exposes the gap between publishing open weights and delivering a usable execution system: Mini can run in Nex's sample two-H100 deployment, Max's reference setup calls for 16 H200s across two nodes, and Pro's weights remain marked as coming soon. Nex-AGI has introduced https://x.com/NexEcosystem/status/2097341149405151287?ref=runtimewire three agent models spanning multimodal computer use and text-based reasoning, but the Pro weights remain marked as coming soon. Nex says the family targets computer use, browsing, reasoning and longer-running agent tasks. The company has placed all three tiers under one release, although the Pro weights have yet to appear in its model repository. The family comprises a 35B Mini, a 397B Pro and a 1.6T Max. Nex-AGI describes Mini and Pro as multimodal models for operating computers and browsers. Max drops visual input and concentrates the largest model on reasoning, coding, scientific work and agent workflows. The people behind the models identify collectively as the Nex-AGI Team. Public materials describe Nex-AGI https://nex-agi.com/?ref=runtimewire as a collaboration involving the Shanghai Innovation Institute https://www.sii.edu.cn/ s3/2026/0607/c27a1089/page.psp?ref=runtimewire , Shanghai Qiji Zhifeng, Mosi Intelligence and Kuafu Technology. The available records do not identify a conventional standalone startup, named founder or chief executive. Nex is associated with the Shanghai institute, though its precise legal structure and headquarters are not publicly established. The N1 technical paper https://arxiv.org/abs/2512.04987?ref=runtimewire , dated December 4, 2025, credits a large research team. The project's public materials present Nex as a technical collaboration rather than centering it on an individual founder. That institutional structure extends beyond model weights. Nex's earlier open-source stack https://github.com/sii-research/NEX?ref=runtimewire covered models, datasets, agent frameworks, reinforcement-learning tools and inference infrastructure. One family, three very different deployments Nex-N2.5 Mini https://huggingface.co/nex-agi/Nex-N2.5-mini?ref=runtimewire is the smallest member of the family, although small remains relative in frontier AI. Mini accepts images and text and is intended for visual computer use, browsing and tasks that require feedback from an interface. Nex-N2.5 Pro https://huggingface.co/nex-agi/Nex-N2.5-Pro?ref=runtimewire is the middle tier, pairing a 397B architecture with multimodal operation. As of September 8, 2026, its model card remained marked as coming soon rather than offering downloadable weight shards. Nex-N2.5 Max https://huggingface.co/nex-agi/Nex-N2.5-Max?ref=runtimewire sits at the other end of the range. Its 1.6T total parameter count makes self-hosting a substantial infrastructure commitment, even though the mixture-of-experts design activates only part of the model for each token. The parameter totals need context. The N2.5 repository https://github.com/nex-agi/Nex-N2.5?ref=runtimewire lists 3B active parameters for Mini, 17B for Pro and 49B for Max. Mixture-of-experts models activate only part of their total parameter count for each token. Nex's materials describe Mini and Pro as post-trained from Qwen3.5 variants and Max as based on DeepSeek-V4-Pro /models/azure/deepseek-v4-pro -Base. N2.5 represents a substantial post-training and systems project, while the underlying base architectures come from other model developers. Nex's work centers on teaching models to plan, use tools, observe an interface, check results and continue through tasks involving many actions. That focus explains why the family spans a smaller visual model, a much larger multimodal tier and a trillion-parameter text system instead of offering one checkpoint for every workload. Nex wants benchmarks to measure sustained work In its N2.5 benchmark announcement https://x.com/NexEcosystem/status/2097341159102443804?ref=runtimewire , Nex reports 50.2 on AutomationBench v1.0.6 for Max and 56.4 on OSWorld-2 for Pro. The figures have not been independently replicated in the reviewed materials. Nex compared Pro's OSWorld-2 result with 46.7 for Qwen3.8-Max /models/qwen/qwen3.8-max . The company has not published enough methodology in the reviewed materials to establish direct comparability. Nex said Pro completed more than 468 actions in Pokemon Platinum in a long-range computer-use demonstration https://x.com/NexEcosystem/status/2097341162210369621?ref=runtimewire . According to the company, the model used visual feedback and standard controls to navigate routes, handle encounters, choose moves, switch characters and manage health. A separate Blender demonstration https://x.com/NexEcosystem/status/2097341172465443175?ref=runtimewire shows N2.5 modifying an existing scene, inspecting the rendered output and revising its presentation. The company also showed the model generating an animated HTML game https://x.com/NexEcosystem/status/2097341169336451579?ref=runtimewire with interactive controls. Another demonstration focuses on visual context https://x.com/NexEcosystem/status/2097341166052364793?ref=runtimewire accumulated from user-recorded screen activity. Together, the demonstrations cover Blender manipulation, interactive HTML and game generation, visual feedback loops and long-horizon computer use. Errors can compound as an agent clicks through interfaces, edits files and runs generated code, so a polished first response says little about whether the system can recover halfway through a long task. Open agent models now compete with execution stacks In materials for the earlier N2 generation, Nex-AGI describes Agentic Thinking https://github.com/nex-agi/Nex-N2?ref=runtimewire as a closed loop connecting requirement understanding, task planning, code implementation, environmental feedback, evaluation, debugging and continued iteration. Max gives the new family a larger text-reasoning tier alongside the two multimodal models. Open weights still leave operators responsible for serving, evaluation, orchestration and accelerator capacity. The gap is especially visible across N2.5: Mini is the smaller and more attainable visual model, while Max's trillion-parameter scale creates a much heavier serving burden. The N2.5 Pro model card's coming-soon status adds a separate availability constraint despite its place in the announcement. Competition is also moving toward systems that execute and evaluate work instead of standalone checkpoints. Computer-agent company Simular reported a 69.9% OSWorld result and a $21.5 million raise https://www.simular.ai/articles/simular-raises-21-5m-to-build-autonomous-computer-agents?ref=runtimewire in December 2025. Browserbase's Stagehand https://docs.browserbase.com/welcome/getting-started?ref=runtimewire , a browser-automation SDK, combines natural-language instructions with deterministic Playwright-style actions. According to OpenAI's AgentKit announcement https://openai.com/index/introducing-agentkit/?ref=runtimewire , AgentKit is a set of tools for developers and enterprises to build, deploy and optimize agents. OpenAI has separately reported 38.1% on OSWorld https://openai.com/index/new-tools-for-building-agents/?ref=runtimewire for its computer-use system and recommends human oversight, an indication of how far reliable interface operation remains from a solved problem. N2.5 packages three tiers under one announcement, with sharply different capability and infrastructure targets. Mini offers the most attainable route to visual execution. Pro targets heavier computer-use workloads, though its model card remains marked as coming soon. Max supplies a large text-reasoning model for operators prepared to absorb its hardware demands.