Nex-N2.6

Nex-N2.6

Handles computer use, coding and long tool-driven tasks as a multimodal agentic model built on GLM-5.3-Flash, in FP8 weights of about 321B parameters. For teams that can self-host a large model.

Made by nex-agi on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Nex-AGI built Nex-N2.6 on top of GLM-5.3-Flash as a multimodal agent model. It reads images and text and answers in text. Its aims are operating a computer, writing code and handling long jobs where tools get called. This repository holds FP8 weights of about 321B parameters.

Who it suits: It suits teams building agents that operate a computer or call tools, and groups that can host a huge model themselves.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Software development | Developers | Run it as the model behind a coding or tool-using agent | An agent framework that sends the tools the model may call
AI research | Researchers | Compare agentic results on tests such as OSWorld | Checking the reported scores against their own test setup
Infrastructure | ML platform teams | Self-host the weights with an FP8-compatible runtime | Serving the model through an SGLang setup

People without hardware for about 321 GB of weights can skip the download. The card also points to an API on OpenRouter.

Where to use it:

• A self-hosted server: any runtime that can serve GLM-5.3-Flash, such as SGLang
• An OpenAI-compatible server: the card's example listens on port 8000
• Agent frameworks: the client or framework runs the tool calls
• OpenRouter: the card lists an API there
• Windows, macOS or Linux: download the files with Python

Title: Jobs it does
Use case | Who | What they get back
Computer use | Agent builders | A model tuned for tasks on a computer, scored on OSWorld
Coding | Developers | Code written on request, such as a short Python function
Tool-using tasks | Agent teams | Tool calls that stay within the tools declared in the request

How to set it up: You need these first:

• Python and pip on Windows, macOS or Linux, to download the files
• Storage for about 321 GB of weights, or an estimated 161 GB at 4-bit
• A serving runtime with GLM-5.3-Flash support and an FP8-compatible setup for your hardware

1. Install the Hugging Face hub package.
2. Download the model files.
3. Follow the SGLang deployment guide the card points to, choosing an FP8-compatible setup.
4. Point the model path at the downloaded folder and set the served model name.
5. Keep the tokenizer, processor and chat template that ship with the weights.

pip install -U huggingface_hub
hf download nex-agi/Nex-N2.6

The served model name is set with this flag.

--served-model-name Nex-N2.6

How to use it: Once the OpenAI-compatible server listens on port 8000, send it chat requests with the model name Nex-N2.6. A few request settings matter:

• reasoning_effort: low, high or max, with max as the default
• stream: set it to true to get the reply as it is written
• max_tokens: caps the length of the reply
• Images: use the image content blocks your serving runtime supports
• Tools: supply the available tools with each request, and your client runs the calls and returns the results

Example: You send a chat request asking, "Write me a Python function that tests whether some text reads the same backward," with reasoning_effort set to max. The reply streams back with the function.

For newcomers and experts: If you are new to this, know that it is a huge model meant for servers, not a laptop. A small first step is to read the card and decide whether to use the hosted API on OpenRouter or to host the weights yourself.

If you work in the field, this release follows Nex-N2.5-Pro and aims at less repetitive reasoning and closer adherence to the tools in the current environment. The team reports reasoning repetition falling from 16.0% to 1.6% in its internal replay evaluation, and 100% tool-list adherence among responses that make tool calls. The team says its scores depend on the evaluation setup and refer to the model overall, not to this FP8 copy on its own. On Terminal-Bench 2.1 and AutomationBench the table shows it below GLM-5.3-Flash.

Download the FP8 weights
Serve them with an FP8-capable runtime
Send OpenAI-style chat requests on port 8000
Supply tools, your client runs the calls

Good to know: The team reports these OSWorld scores for this release and its comparison models.

Title: OSWorld score reported by the team
Nex-N2.6: 83.0 points
Nex-N2.5-Pro: 82.2 points
GLM-5.3-Flash: 78.0 points
DeepSeek-V4.1-Flash: 71.7 points

• License: apache-2.0
• Cost to run: the card points to an API on OpenRouter, and self-hosting needs hardware for about 321 GB of weights, with no prices named
• Data it reaches: whatever you send in a request, and your own client decides what files or services tool calls touch
• Not for: a guarantee of these scores on your setup, since they depend on the evaluation setup and settings

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 321B
• Weights: about 321 GB in F8_E4M3; about 161 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 191
• Likes: 60
• License: apache-2.0
• Quantized versions: 6
• Tags: transformers, safetensors, glm5_next, image-text-to-text, conversational, fp8

Links

• Model: https://huggingface.co/nex-agi/Nex-N2.6
• Files: https://huggingface.co/nex-agi/Nex-N2.6/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nex-agi/Nex-N2.6

Then follow the usage in its model card. Nothing in this kit runs by itself.