Nex-AGI’s readable Nex-N2.6 model card describes FP8 weights, instructions for running the model on your own hardware and company-reported evaluation results. It identifies GLM-5.3-Flash as the base model and puts the model at approximately 321 billion parameters. Nex-AGI’s model card

Treat the scores as results Nex-AGI reports for its evaluation setups, not as a guarantee of performance in every workload.

What Nex-AGI reports in its evaluations

Nex-AGI reports results for Nex-N2.6, Nex-N2.5-Pro and GLM-5.3-Flash across three evaluation suites. Nex-N2.6 scores higher than Nex-N2.5-Pro on all three listed metrics, while GLM-5.3-Flash scores slightly higher on Terminal-Bench 2.1 and AutomationBench. Nex-AGI’s model card

EvaluationNex-N2.6Nex-N2.5-ProGLM-5.3-Flash
OSWorld83.082.278.0
Terminal-Bench 2.184.182.784.3
AutomationBench45.744.248.8

These are Nex-AGI-reported, model-level scores. The company says results depend on the evaluation harness and settings; the scores are not a separate evaluation of the FP8 export.

The card also describes an internal known-issue replay evaluation: it fixes the preceding conversation and tool outputs, then evaluates the model’s next response. In that test, Nex-AGI reports a reasoning-repetition rate of 1.6% for Nex-N2.6, compared with 16.0% for Nex-N2.5-Pro. That result applies to this replay test, not to every conversation or task. Nex-AGI’s model card

What self-hosting requires

Nex-AGI’s instructions call for an inference runtime that supports GLM-5.3-Flash and hardware compatible with the model’s FP8 weights. The card says to keep the supplied tokenizer, processor and chat template when setting up the model, and includes an example of an OpenAI-compatible local server. Nex-AGI’s model card

FP8 refers to an 8-bit floating-point format for the model weights. The card describes Nex-N2.6 as approximately 321 billion parameters, so self-hosting means working with a large model; the published instructions identify compatible software and hardware, not a universal setup for every machine.

OrcaRouter’s reported API rates

On November 10, 2026, OrcaRouter reported that Nex-N2.6 was live on Nex-AGI’s commercial API and listed its rates at $0.12 per million input tokens and $0.40 per million output tokens. Its comparison put Nex-N2.5 Pro at $0.075 per million input tokens and $0.25 per million output tokens. By those listed figures, each Nex-N2.6 rate is 60% higher than the corresponding Nex-N2.5 Pro rate.

Input and output tokens are billed separately, so the cost depends on how many of each a request uses. OrcaRouter’s figures are a reported listing comparison, rather than a US-specific price quote.