Three Nex-N2.5 models to act, code, or orchestrate

Nex-AGI introduces Nex-N2.5 Mini, Pro, and Max, three open agentic models designed for computer use, coding, and long-running tasks.

An agent opens Blender, edits a scene, reviews the result, and then corrects its work. Another completes an expense reimbursement request, moves between several screens, and checks that the transaction has been recorded. Nex-AGI uses these examples to introduce Nex-N2.5, a new family of models designed to work inside software rather than merely respond in a chat window.

The lineup includes three models with different roles. Mini and Pro accept text and images. They can analyze a screenshot, locate a control within an interface, and propose the next action. Max only accepts text and is geared more toward complex reasoning, coding, and the orchestration of long tool chains.

Mini has 35 billion parameters. Its configuration uses the multimodal Qwen3.5 MoE architecture and specifies a 262,144-token context window. Its BF16 files take up about 70 GB on Hugging Face. The model contains 256 experts, eight of which are activated for each token. This selective activation reduces the amount of computation performed at each step compared with a dense model of equivalent total size.

Pro is announced with 397 billion parameters. It extends the multimodal foundation used by Nex-N2-Pro, which is itself based on Qwen3.5-397B-A17B. The A17B suffix indicates that only a portion of the network, representing roughly 17 billion parameters, is activated for each token. This architecture seeks a balance between the model’s total capacity and the cost of running it.

Max changes both scale and foundation. Nex-AGI describes it as a 1.6-trillion-parameter MoE model built from DeepSeek-V4-Pro-Base. The total parameter count does not represent the amount of computation triggered for each word, since the mixture-of-experts system activates only part of the network. Nex-AGI does not provide a single figure in its launch materials for the exact number of parameters activated per token.

The three versions are therefore not a simple progression from a small model to a very large one. Mini and Pro include a visual component, while Max does not. Max cannot directly inspect a browser window or the state of an application. To interact with a computer, it must receive a text description of the environment or work alongside another system responsible for perception.

This distinction qualifies the claim that the family can use computers. Computer use with Mini and Pro follows an observation-based loop. The agent receives an image of the screen, interprets the state of the application, selects an action, and then examines the next image to verify its effect. It can adjust its strategy when the expected button does not appear, a dialog box blocks the interface, or the result does not match its objective.

The model does not control the machine by itself. An external program must capture the screen, convert its responses into clicks, keystrokes, or commands, and then return the updated state. This layer also determines which applications the agent can access, how long it may operate, what permissions it receives, and when it must stop.

Nex-AGI says it expanded the environments used during post-training. Its examples include browsers, 3D creation software, computer-aided design tools, administrative tasks, and some PC games. The goal is to sustain activity beyond a few isolated operations and teach the model to verify the consequences of its decisions.

The demonstrations show this continuity, but they do not provide real-world failure rates for each application. The company does not disclose how many attempts were required, which sequences were selected, or how many commands were prepared in advance. They illustrate the intended use cases without serving as an independent measure of reliability.

The ability to self-correct also does not guarantee that a task will eventually succeed. A perception error can lead to an incorrect action, which may then become the starting point for several further decisions. During a long operation, the agent must also retain the original constraints, recognize what has already been completed, and avoid repeating an action whose effect is not immediately visible.

Max targets a different kind of continuity. The model is designed for large codebases, research involving many steps, and agents that call multiple services. Its configuration declares support for up to 1,048,576 positions. The official deployment recipe, however, limits the server to 262,144 tokens. The length specified in the configuration file therefore does not automatically match the one used in the recommended setup.

Nex-N2.5 offers three reasoning settings. `none` generates a direct response, `medium` lets the model determine how much effort is needed, and `high` systematically activates a reasoning phase. The intermediate setting is selected by default. The family also supports function calls, allowing an agent to choose a tool and prepare its parameters in a format that an application can process.

Integration uses requests compatible with OpenAI’s Chat Completions format. Nex-AGI recommends a customized version of SGLang distributed in the `nexagi/sglang:v0.5.18-nex-patch` Docker image. Mini and Pro use a Qwen3 reasoning parser, while Max uses the DeepSeek-R1 parser. All three rely on the Qwen3 Coder parser for tool calls.

The required hardware is a reminder that “small” is relative. The published recipe calls for two H100 GPUs for Mini and eight H100s for Pro. Max requires sixteen H200s distributed across two machines. These are Nex-AGI’s recommended configurations, not necessarily the only possible setups, but they clearly establish the target for official deployment.

The company developed an inference engine called NexRT for Pro. It runs on eight H100 or H200 GPUs and combines SGLang for context prefill with several speculative generation techniques. Nex-AGI notes that throughput depends on the acceptance rate of tokens proposed in advance. Performance measured in an optimized demonstration therefore cannot be applied directly to every request or