Nex-N2.6: API Access, Pricing, Benchmarks and Self-Hosting

Last updated: November 14, 2026

Nex-N2.6 is Nex-AGI’s new open-weight agentic model, released November 10, 2026. It is post-trained on GLM-5.3-Flash, has about 321 billion parameters in FP8, accepts text and images, and targets computer use, coding and tool-heavy tasks. You can call it on OpenRouter or Nex-AGI’s MOSS API, or self-host it under Apache 2.0.

This guide covers what changed from Nex-N2.5-Pro, how to call the model, what it costs, how to run the weights yourself, and the limits to plan around. All figures come from the Nex-N2.6 model card, the OpenRouter listing and the MOSS API model page.

Contents

What is new in Nex-N2.6
Benchmarks compared
Where to use it and what it costs
How to call the API
How to self-host the weights
Limitations
Frequently asked questions

What is new in Nex-N2.6

Nex-N2.6 follows Nex-N2.5-Pro, which OpenRouter lists as a September 2026 model with a Qwen3 tokenizer. The biggest change is the base: Nex-N2.6 is built on GLM-5.3-Flash, and Nex-AGI says the release “focuses on reducing repetitive reasoning and improving adherence to the tools available in the current environment.”

The model card highlights three results, all from Nex-AGI’s own evaluations:

Less repetitive reasoning. The reasoning repetition rate fell from 16.0% on Nex-N2.5-Pro to 1.6%, which Nex-AGI describes as a 90% relative reduction.
Tool-list adherence. Every tool call in Nex-AGI’s test used a tool declared in the request (100% valid tool selection), up from 82.2% for Nex-N2.5-Pro.
Computer use. OSWorld rose to 83.0, from 82.2 for Nex-N2.5-Pro and 78.0 for GLM-5.3-Flash.

The weights are published on Hugging Face and mirrored in the sii-nex/Nex-N2.6 GitHub repository, both under Apache 2.0.

Benchmarks compared

Nex-AGI notes that scores depend on the evaluation harness and settings, and that they are model-level results rather than a separate test of the FP8 export.

Benchmark	GLM-5.3-Flash	DeepSeek-V4.1-Flash	Nex-N2.5-Pro	Nex-N2.6
OSWorld	78.0	71.7	82.2	83.0
Terminal-Bench 2.1	84.3	82.7	82.7	84.1
AutomationBench	48.8	54.8	44.2	45.7
Reasoning repetition rate (lower is better)	6.0%	Not reported	16.0%	1.6%
Valid tool selection rate, overall	85.3%	Not reported	82.2%	100.0%

The pattern is clear: Nex-N2.6 leads on computer use and tool discipline, roughly matches its base model on terminal coding, and trails DeepSeek-V4.1-Flash and GLM-5.3-Flash on AutomationBench. For background on those two models, see our DeepSeek V4.1 Flash guide and the Vercel Token Pass guide, which covers GLM 5.3 Flash access.

Where to use it and what it costs

Option	Model ID	Price	Context and output
OpenRouter (served by Nex AGI, FP8)	`nex-agi/nex-n2.6`	$0.12 input, $0.40 output, $0.024 cache read per million tokens (20% discount shown)	262,144-token context; up to 235,929 completion tokens
MOSS API (Nex-AGI)	`nex-n2.6`, snapshot `nex-n2.6-2026-11-05`	Free for a limited time	Default output 8,192 tokens; maximum 131,072
Self-hosted	Your served name, for example `Nex-N2.6`	Your hardware cost; Apache 2.0	Depends on your runtime settings

For comparison, OpenRouter lists Nex-N2.5-Pro at $0.075 input and $0.25 output per million tokens, so Nex-N2.6 costs more per token. Z.ai’s own GLM 5.3 Flash endpoint on OpenRouter is listed at $0.15 input and $0.50 output with a 1,048,576-token context, which is four times the Nex-N2.6 window there.

OpenRouter lists support for tools, tool_choice (none, auto, required or a named function), structured outputs through response_format, reasoning_effort, logprobs and implicit caching. Structured outputs through response_format are new compared with the Nex-N2.5-Pro listing. If you route between several open models, our OpenRouter Auto Code guide explains the router options.

How to call the API

The MOSS API offers Chat Completions (/v1/chat/completions), Responses (/v1/responses) and Anthropic-style Messages (/v1/messages), plus input token counting. To start:

Sign in to the MOSS API console and create an API key.
Point an OpenAI SDK at base_url="https://api.mosi.cn/v1" and send the key as a Bearer token. For Messages and token counting, send it as x-api-key with anthropic-version: 2023-06-01.
Set model to nex-n2.6, or to nex-n2.6-2026-11-05 when you need a fixed snapshot for reproducible runs.
Pass the full conversation history with each request. Responses calls are stateless.
Declare your tools in every request. Your client executes the calls and returns the results; function calling works in Chat Completions and Messages.
Set max_tokens (or max_output_tokens in Responses) above the 8,192 default for long agent turns, up to 131,072.

Images work in all three formats. Video (MP4 or WebM through video_url) works in Chat Completions and Responses but not Messages. MOSS does not accept file uploads or file_id references for this model, media URLs must be publicly reachable, and the encoded request body must stay under 10 MiB. Each user can run up to 10 concurrent requests. On OpenRouter, use the ID nex-agi/nex-n2.6 with the usual OpenRouter endpoint.

How to self-host the weights

The Hugging Face repository holds FP8 weights split into 129 safetensors files, about 328 GB in total. The model card gives this path:

Install an inference runtime with GLM-5.3-Flash support. Nex-AGI points to SGLang’s GLM-5.3-Flash deployment guide; our SGLang 0.5.23 guide covers recent SGLang changes.
Choose an FP8-compatible configuration for your GPUs.
Download the Nex-N2.6 repository and replace the model path in the SGLang command with that directory.
Set --served-model-name Nex-N2.6 and keep the tokenizer, processor and chat template shipped with the weights.
Call the OpenAI-compatible server on port 8000 at /v1/chat/completions.

The bundled chat template accepts reasoning_effort values low, high and max, and defaults to max. Lower the effort for routine steps if latency matters. For other large open models you can run on your own hardware, see our Edge0-145B guide.

Limitations

Self-reported results. Every benchmark and reliability figure on the model card comes from Nex-AGI’s own evaluation.
Text output only. The model reads text and images (and video on MOSS) but returns text only. MOSS lists no audio, speech or image generation for it.
Shorter hosted context than the base. OpenRouter serves Nex-N2.6 at 262,144 tokens, while GLM 5.3 Flash endpoints there list up to 1,048,576.
Few hosts. At the time of writing, OpenRouter lists one provider, Nex AGI, so plan a fallback model for production agents.
Temporary pricing. The MOSS free period and the OpenRouter discount have no published end date.

Frequently asked questions

What is Nex-N2.6?

Nex-N2.6 is a multimodal agentic model from Nex-AGI, post-trained on Z.ai’s GLM-5.3-Flash and aimed at computer use, coding and long tool-using tasks. The FP8 weights have about 321 billion parameters and are published on Hugging Face under the Apache 2.0 license.

How much does Nex-N2.6 cost?

On OpenRouter it costs $0.12 per million input tokens, $0.40 per million output tokens and $0.024 per million cached input tokens, shown with a 20% discount. Nex-AGI’s MOSS API lists the model as free for a limited time.

What is the Nex-N2.6 context window?

OpenRouter lists a 262,144-token context window with up to 235,929 completion tokens. On the MOSS API the default output budget is 8,192 tokens and the maximum is 131,072 tokens.

Is Nex-N2.6 better than GLM-5.3-Flash?

In Nex-AGI’s own results it scores 83.0 on OSWorld against 78.0 for GLM-5.3-Flash, but slightly lower on Terminal-Bench 2.1 (84.1 against 84.3) and AutomationBench (45.7 against 48.8). These figures are self-reported.

Can I run Nex-N2.6 locally?

Yes, if you have server-class hardware. The FP8 repository holds about 328 GB of weight files. Nex-AGI recommends SGLang’s GLM-5.3-Flash deployment guide with an FP8-compatible configuration and the tokenizer and chat template shipped with the weights.

Nex-N2.6: Power-User Guide to the API, Pricing and Self-Hosting