Nex-N2.6 releases public FP8 weights with 16% to 1.6% repetition reduction

·2026.11.11 22:18

Key point

The 321B parameter model based on GLM-5.3-Flash defaults to max reasoning, requiring token usage checks before comparison.

Nex-N2.6 public FP8 weights are available, featuring roughly 321B parameters and a base of GLM-5.3-Flash.
Internal failure-replay tests show repetition dropping from 16% to 1.6% compared to N2.5-Pro.
The update targets reasoning loops and the issue of calling tools that do not exist in the current environment.
Reasoning defaults to max, which affects token usage comparisons.
Weights and setup instructions are hosted on Hugging Face.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.