AIAsia / North America

Axion’s new 35B download is a community reproduction with its own results

A November 4 project discussion announced reproduced Axion weights. The official repository now links the community model while distinguishing it from an official checkpoint.

iNewsChannel IntelligenceAI-assisted reporting · Source-linked analysis

Source published: Nov 4, 2026, 4:15 AM UTC
Published: Nov 5, 2026

1 min · Reading time

Original Axion chart compares the paper’s 35B and 122B models with other systems. — Axion 1.0 project authors | Apache-2.0 · Apache License 2.0 — associated documentation · Project-reported paper results, not measurements of the community reproduction; the article explains the difference.

Added after our earlier coverage. The original source date is shown below; this is not a new announcement today.

What happened

The Hugging Face release contains a 35B mixture-of-experts model with about 3B active parameters, under Apache 2.0. Its model card reports SEAL-0 at 60.4 versus the paper’s 62.2 and FrontierScience Research at 55.0 versus 56.7. These are publisher-reported scores.

iNC / Analysis

Why it matters

A reproducible training recipe and downloadable weights are different deliverables. This checkpoint gives developers a concrete artifact to test, but the paper’s stronger numbers cannot be assigned to it. Tool-use tests also require the execution harness, not just the model.