Nex-N2.6 opens 321B agent weights, with reliability caveats
Nex-N2.6 is available as public FP8 weights and through an API. Nex AGI built the 321 billion parameter coding and computer use model on GLM-5.3-Flash, targeting repetitive reasoning and incorrect tool selection.
Availability was checked on November 11, 2026. OpenRouter dates its listing to November 10, while the repository history contains uploads on November 10 and edits on November 11. Those commit dates do not establish the exact hour the weights became public.
The reliability claim needs context
Nex reports repetition falling from 16% with Nex-N2.5-Pro to 1.6%, plus 100% adherence to declared tools, in internal checks. These replay the next response after a fixed conversation. They do not establish that an entire autonomous job finishes correctly.
The published benchmarks are not a separate evaluation of this FP8 export. ByteForward has not independently tested the release.
That distinction matters in practice. Choosing a tool that exists is only one part of completing a job. An agent can select the right browser tool but click the wrong control, reuse stale information or stop before checking its work. A strong next response still needs to survive the rest of the task.
API pricing and access
The OpenRouter service provides a quick way to evaluate the model without operating the weights. Its current listing shows the following configuration and prices in US dollars.
| OpenRouter listing | November 11 snapshot |
|---|---|
| Model identifier | nex-agi/nex-n2.6 |
| Input and output | Text and images in, text out |
| Context window | 262,144 tokens |
| Input price per million tokens | $0.12 |
| Output price per million tokens | $0.40 |
| Cached input per million tokens | $0.024 |
The prices are marked as a 20% discount. The page does not give an end date in the inspected listing, so budget against current checkout and API pricing rather than assuming the discount is permanent. It also says this model is unavailable through EU in-region routing.
For an illustrative batch using 10 million uncached input tokens and one million output tokens, the listed token rates total $1.60. That arithmetic excludes any separately billed tools or infrastructure. Compare cost per successfully completed task as well as price per token: retries and longer answers can change the total.
Self hosting is a large deployment
The official file tree totals about 328 GB and lists 129 weight shards. Download size is not a complete GPU memory requirement. Serving also needs room for runtime allocations, request state and concurrency, so this should not be treated as a typical laptop model.
Use a runtime supporting GLM-5.3-Flash and FP8, preserving the supplied tokenizer, processor and chat template. Reasoning supports low, high and max. The default is max. Record that setting in comparisons.
The repository labels the release Apache 2.0. Review the applicable notices and the complete deployment stack before redistribution. A model license does not settle the permissions for the repositories, accounts or tools an agent will use.
For a much smaller local agent option, our LFM3 guide explains another deployment pitfall: a file can load successfully while an incompatible runtime produces incorrect output. It addresses a different model size and workload, not an equivalent replacement.
Test the failure that costs you time
Start with a task where your current agent gets stuck, such as repeatedly proposing the same fix after a failed test. Keep the starting repository, tool descriptions, permissions and stopping rules identical. Log repeated actions, invalid tool selections, completed work, elapsed time and total cost.
Add a separate case where a tempting tool is deliberately absent. Then check whether the agent asks for help, uses an available alternative or invents a call. Use a disposable workspace and mock accounts for these trials. This is a suggested evaluation plan, not a ByteForward test result.
The useful decision is whether this release reduces the failure that wastes your time without introducing a different one. A replay benchmark can help identify what to test. Your own complete workflow determines whether the change is worth adopting.
AI generated editorial illustration by ByteForward. NEX wordmark sourced from Nex-AGI.
