Nex-AGI Releases Nex-N2.5 Open Agents That Browse, Click, and Code
Nex-AGI open-sourced a three-tier agentic model family under Apache 2.0, topped by a 1.6T-parameter MoE built to run browsers, desktops, and code loops end to end.
- Nex-AGI released Nex-N2.5 agentic model family in mini, Pro, and Max sizes under Apache 2.0.
- Max is a 1.6-trillion-parameter text-only MoE, the lab's first trillion-scale post-training run.
- Models use vision as a feedback loop: observe screen, act, verify, self-correct.
- Pro leads OSWorld-G GUI grounding at 87.4; Max leads BrowseComp at 92.6.
- Mini runs on 2xH100, Pro on 8xH100, Max on 2 nodes of 16xH200.
- Hosted access on OpenRouter; SGLang Docker image with OpenAI-compatible API for self-hosting.
Nex-N2.5 brings open-weight agents to browsers, desktops, and code
Nex-AGI has released Nex-N2.5, a family of agentic models designed for long-horizon work across browsers, desktop interfaces, terminals, and code environments. The weights use the Apache License 2.0, and hosted inference is available through OpenRouter for teams without local GPU capacity.
Three sizes, very different machines
| Tier | Design | Context | Reference deployment |
|---|---|---|---|
| Nex-N2.5-mini | Multimodal, roughly 35B total parameters and 3B active | 262,144 tokens | 2 H100 GPUs, tensor parallelism 2 |
| Nex-N2.5-Pro | Multimodal, tuned for computer use, browsing, and visual grounding | 262,144 tokens | 8 H100 GPUs, tensor parallelism 8 |
| Nex-N2.5-Max | Text-only, 1.6-trillion-parameter Mixture-of-Experts model | 262,144 tokens | 2 nodes with 16 H200 GPUs, expert parallelism 16, and DeepEP routing |
Mixture-of-Experts models route each token through a subset of specialized parameter groups. Sparse activation reduces per-token computation relative to evaluating every parameter, although serving still requires enough memory to hold the weights. Nex-AGI describes Max as its first complete post-training effort at trillion-parameter scale.
Mini contains 256 routed experts, activates eight per token, and adds one shared expert. Its MoE intermediate dimension is 512. That configuration gives developers the family’s smallest self-hosted entry point, although it still requires two H100 GPUs under the published setup.
Screens close the control loop
For mini and Pro, screenshots serve as feedback during browser and desktop tasks. The model inspects the current interface, selects an action, observes the resulting state, and adjusts its next step. That loop lets an agent verify whether a menu opened, a form submitted, or a test passed before continuing.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
