Nex-N2.6

Nex-AGI's agent models for computer use and coding, led by N2.6 on GLM-5.3-Flash (321B, FP8) and the N2.5 mini, Pro and Max line.

In short. Nex-N2.6 is an open AI model built to act as an agent: it can look at a screen, operate a computer or a browser, write and test code, and call tools over long tasks. It is built on GLM-5.3-Flash and is about 321 billion parameters. The same entry covers the earlier Nex-N2.5 models in three sizes, from a 35B mini to a 1.6T Max.

Summary

Nex-AGI, the group behind the SII Nex GitHub organization, published Nex-N2.6 on Hugging Face on 10 November 2026. It is a multimodal agent model built on GLM-5.3-Flash, with about 321B parameters in FP8 weights of 328 GB, aimed at computer use, coding and long tool-using tasks. Compared with Nex-N2.5-Pro, the card says it cuts reasoning repetition from 16.0% to 1.6% and uses only the declared tools in 100% of the responses that make tool calls, both in the team's internal evaluations. The entry also covers the Nex-N2.5 family from September 2026, which has a 35.1B mini, a 396.8B Pro that accepts images and a 1.6T Max that is text-only and built on DeepSeek-V4-Pro-Base.

For N2.6 the card reports 83.0 on OSWorld, against 82.2 for Nex-N2.5-Pro and 78.0 for GLM-5.3-Flash. On Terminal-Bench 2.1 it scores 84.1 against 84.3 for GLM-5.3-Flash, and on AutomationBench 45.7 against 48.8, so it is not ahead of its base on every row. The N2.5 card reports Terminal-Bench 2.1 scores of 73.4 for mini, 82.7 for Pro and 86.1 for Max, against 89.1 for Claude Opus 5 and 88.3 for Kimi-K3, and SWE-Bench Pro scores of 43.8, 61.2 and 65.7, against 79.2 for Claude Opus 5 and 63.3 for Kimi-K3. For computer use, Pro scores 82.2 on OSWorld-Verified against 83.4 for Claude Opus 5 and 84.8 for Kimi-K3.

The N2.5 scores come from Nex-AGI's NexAU coding and NexCUA computer-use harnesses, with public figures for rivals where they exist and its own runs otherwise, at temperature 0.7, top-p 0.95 and top-k 40. The cards give SGLang commands using a patched Docker image, with two H100 GPUs for mini, eight H100 GPUs for Pro and two nodes of eight H200 GPUs for Max. The N2.5 template takes reasoning effort none, medium or high, and the N2.6 template takes low, high or max. The config files set a position limit of 262,144 tokens for mini and Pro and 1,048,576 for N2.6 and Max, but the cards state no context length. Hugging Face tags all four repositories Apache 2.0. In the 30 days to 11 November 2026 it counted about 3,000 downloads for mini, 1,900 for Pro and 450 for Max, and N2.6 was a day old.

Good for

  • Computer-use and browser agents that work from screenshots, with N2.6 or Pro
  • Coding and terminal agents on long tasks, using the tool-calling support in the SGLang setup
  • A multimodal agent on a single node with two H100 GPUs, using the 35B mini
  • Teams that want an Apache 2.0 agent model close to proprietary models on the developer's own tables

Watch out for

  • Most scores are Nex-AGI's own, run in its NexAU and NexCUA harnesses, and the card says the NexCUA project was still to be open-sourced
  • N2.6 does not beat its GLM-5.3-Flash base on Terminal-Bench 2.1 or AutomationBench, and its reliability figures come from an internal replay evaluation, not a public benchmark
  • The weights are large, about 328 GB for N2.6, 407 GB for Pro and 1.65 TB for Max, so only the 70 GB mini fits a small multi-GPU box
  • Nex-N2.5-Max is text-only, and N2.6 needs a runtime that supports GLM-5.3-Flash, so check which repository and serving image you use

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Nex-N2.5-mini variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Nex-N2.5-mini BF16 35.1B 81 GB 40 GB 24 GB Gaming GPU
Nex-N2.6 (FP8) FP8 321B 739 GB 370 GB 222 GB Big-memory desktop
Nex-N2.5-Pro (FP8) FP8 397B 913 GB 456 GB 274 GB Big-memory desktop
Nex-N2.5-Max (FP8) FP8 1.6T 3.68 TB 1.84 TB 1.1 TB GPU server

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Nex-N2.6?
Nex-N2.6 is an open AI model built to act as an agent: it can look at a screen, operate a computer or a browser, write and test code, and call tools over long tasks. It is built on GLM-5.3-Flash and is about 321 billion parameters. The same entry covers the earlier Nex-N2.5 models in three sizes, from a 35B mini to a 1.6T Max.
Can I use Nex-N2.6 commercially?
Yes. Nex-N2.6 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Nex-N2.6 open source?
Nex-N2.6 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Nex-N2.6 on my own computer?
Yes, on a good graphics card. The smallest variant, Nex-N2.5-mini, needs about 24 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was Nex-N2.6 released?
Nex-N2.6 was released by Nex-AGI on November 10, 2026.
How long a context window does Nex-N2.6 have?
Nex-N2.6 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
GLM-5.3-Flash Z.ai (Zhipu AI) 320B 18B active 1M MIT (Permissive) 221 GB Big-memory desktop
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone
DeepSeek V4 DeepSeek 284B – 1.6T 49B active 1M MIT (Permissive) 176 GB Big-memory desktop
Kimi K3 Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster