Skip to content

AI

A 1-Trillion-Parameter Model Just Beat GPT-6 Astra at Reading Powder

Periodic Labs says Neon hits 55.3% on X-ray diffraction analysis, up from 2.7% for its base model. The catch: the exam is the company's own, and the model stays in-house.

Cover photo: HaeB / CC BY-SA 4.0

TL;DR

  1. Periodic Labs' Neon scores 55.3% on its internal X-ray diffraction test; the base model scored 2.7%.
  2. Periodic says Neon beats GPT-6 Astra and Claude Fable 5.1 on FrontierXRD, its own benchmark.
  3. Neon is built on Kimi K2.6; weights and training data remain unreleased.

What happened

Periodic Labs has introduced Neon, a one-trillion-parameter model specialized in analyzing experiments run in its own high-throughput labs.

The target task is X-ray diffraction: you fire X-rays at a powdered material, get a pattern of peaks, and work out which crystalline phases you actually made. With overlapping phases, an experienced scientist can spend hours on it.

On FrontierXRD, an internal evaluation focused on diffraction measurements, Neon scores 55.3%. Its base model scores 2.7%. Periodic says Neon also beats GPT-6 Astra and Claude Fable 5.1 on that same test.

This is not a from-scratch foundation model. Periodic began with Kimi K2.6, an open-weight base of the same one-trillion size, then layered midtraining and reinforcement learning on scientific and lab data. Periodic also says Neon did it with relatively little compute.

Neon is internal for now. Neither the weights nor the training corpus has been released.

Why it matters

The pitch is a loop. Periodic's labs run experiments, the experiments produce data, the data trains scientific AI, and that AI guides the next experiments.

The company says it uses frontier models while they work, and trains its own when they fall short or get too costly.

That is a quiet argument against the "one giant general model does everything" story. Proprietary lab data is the one thing a chatbot cannot scrape, and Periodic is betting it is worth more than raw scale.

Now the fine print, sapiens. The benchmark is Periodic's own, the comparison is vendor-reported, and nobody outside the company has verified it. A 55.3% score also means the model still misses a lot of powder patterns.

A specialist trained on private lab data beat generalists on a test its maker wrote. Useful signal, not a verdict.

The lab side is real, though: Periodic says it has built high-throughput labs in Menlo Park, running 24/7, aimed at new materials such as superconductors and magnets.

What to do with it

  • Read the benchmark, not the headline.FrontierXRD is internal, so treat the win over GPT-6 Astra and Claude Fable 5.1 as a claim until independent tests appear.
  • Steal the pattern.If your team owns unusual domain data, a tuned open-weight base can be cheaper than waiting for the next frontier release.
  • Watch for a release.Weights and training data are unreleased, so check whether Periodic opens anything up before you plan around Neon.
  • Track materials AI.If you work in chemistry, batteries or magnets, lab-embedded models are the part of this story to follow.

Keep up, sapiens. The powder is already being read.