Nex AGI Releases Nex-N2.6 Multimodal Agentic Model

Reliability in tool use is another standout metric. Nex-N2.6 demonstrated 100% tool-list adherence in internal evaluations, ensuring that every tool call made was valid and within the provided schema. This is supported by an overall non-repetition rate of 97.8% and a 100% valid tool selection rate when tools are present.

• OSWorld Score: 83.0 (vs DeepSeek V4.1 Flash: 71.7)
• Terminal-Bench 2.1: 84.1
• Reasoning Repetition Reduction: 90% relative decrease
• Tool Call Adherence: 100%

Agentic Capabilities: From Pixels to 3D

To demonstrate the practical utility of Nex-N2.6, Nex AGI showcased an end-to-end agentic 3D workflow. In a live demo, the model was tasked with building a complex Blender model starting from nothing but a single photograph. The model successfully analyzed the image, matched materials and lighting, and eventually rendered a fully animated showcase.

This level of cross-modal reasoning—translating 2D visual data into 3D spatial structures and procedural instructions—highlights the model's readiness for professional creative pipelines and automated engineering tasks.

• End-to-end 3D modeling from single-image input
• Automated material and lighting matching
• Procedural animation and rendering orchestration

API Pricing

Nex-N2.6 is highly accessible via both managed APIs and open-source deployment. For developers looking for low-latency, managed access, the model is live on OpenRouter and the MOSS API platform. The pricing is designed to be competitive for high-volume agentic workflows that require large context windows.

With a 262k context window, the model is ideal for RAG-heavy applications and long-form document analysis. The cost structure remains extremely lean, even for a 321B parameter model, making it viable for production-scale deployment.

• Platform Availability: OpenRouter, MOSS API
• Context Window: 262,000 tokens
• Quantization: FP8 available

Use Cases

Nex-N2.6 is optimized for several high-impact domains. Its primary strength lies in 'Agentic Coding,' where it can not only write code but also navigate a terminal, run tests, and debug errors in real-time. Its multimodal nature makes it a prime candidate for 'Computer Use' agents that automate UI-based workflows.

Other key applications include complex RAG (Retrieval-Augmented Generation) involving video and image datasets, long-horizon task planning in robotics simulations, and automated 3D asset generation for game development and VFX.

• Autonomous Software Engineering & Debugging
• GUI-based RPA (Robotic Process Automation)
• Multimodal RAG (Text, Image, Video)
• 3D Asset Creation and Scene Orchestration

Getting Started

Developers can immediately start experimenting with Nex-N2.6 by accessing it through the MOSS API platform or via OpenRouter. For those preferring local deployment, the FP8-quantized weights are available under the Apache 2.0 license, allowing for full control over the inference stack.

The model is ready for integration into existing agent frameworks like LangChain or AutoGPT, thanks to its perfect tool-calling adherence and standardized output formats.

• API Access: MOSS API & OpenRouter
• Local Deployment: FP8 weights via Shanghai Innovation Institute
• License: Apache 2.0

API Pricing — Input: $0.12 per million tokens / Output: $0.40 per million tokens / Context: 262k

Sources