Periodic Labs trains a trillion-parameter model using 1,300 H200 GPUs for scientific experiments.

Periodic Labs has released the scientific model Periodic Neon: starting from the open-weight, trillion-parameter model Kimi K2.6, it underwent intermediate training and reinforcement learning using material data generated by the company’s lab, then was deployed back to the lab to analyze X-ray diffraction results. On the challenging internal benchmark FrontierXRD, success rates improved from 2.7% for the base model to 55.3%; Periodic claims it outperforms GPT-6 Astra and Claude Fable 5.1 on this specific task while being more cost-efficient. What truly matters is not the simplistic comparison of “1,300 H200s beating 100,000 GPUs,” but rather the company’s establishment of a closed loop: experimental labs continuously generate new data unavailable on the internet, the model learns from it, and then determines what to synthesize and test in the next round. However, all major achievements so far come from internal evaluations relying on LLM judges, with no peer review or independent replication yet; Neon inherits the expensive pretraining of Kimi K2.6, and as of this retrieval, no public download link for Neon’s weights has been found.

Article author, source: Periodic Labs

Periodic Labs is searching for better superconductors and magnetic materials. Experimenters can set target materials, select raw ingredients and temperatures, but the resulting powder may not be the expected product—it might contain the target crystal phase, unreacted starting materials, and unintended byproducts.

Researchers typically use powder X-ray diffraction, or XRD, to determine the composition of a sample. When X-rays strike a crystal, they produce a set of peaks; different crystal structures generate distinct patterns, similar to a material’s “fingerprint.”

However, real samples often contain multiple substances whose peaks overlap. A mathematically well-fitted solution may still be chemically unreasonable. Analysts must also consider the raw materials, temperature, atmosphere, whether the sample came into contact with moisture, previous experiments, crystallographic databases, thermodynamic calculations, and relevant publications to determine which phases are truly present and their respective proportions.

Periodic says that complex samples may require materials scientists to analyze for hours. In its evaluations, the accepted answers averaged five phases. This is precisely what Neon automates: reading diffraction patterns and full experimental contexts, invoking databases and scientific software, proposing candidate phases, and iteratively testing its interpretations.Research articlestates that the model has been deployed in the company’s high-throughput laboratory.

Neon is not a newly pre-trained foundational model from scratch. Periodic starts from the open-weight Kimi K2.6, undergoes intermediate training with a scientific corpus composed of academic papers, code, and proprietary experimental data, and then fine-tunes using laboratory data through reinforcement learning.

On the company's most challenging internal evaluation, FrontierXRD, the original Kimi K2.6 achieved a success rate of 2.7%, while Neon reached 55.3%, approximately a 20-fold improvement. The evaluation included 134 complex samples from Periodic Labs.

The company also tested 198 XRD measurements from retained chemical systems: these systems and larger systems containing them were not included in the training data, although smaller sub-systems may have appeared. Neon again outperformed the leading models on this test set, though Periodic explicitly noted that this set is easier than FrontierXRD.

(The original includes a performance-cost scatter plot, a reinforcement learning computational scaling curve, and a generalization results plot on the chemical system.)

The claim of "exceeding GPT-6 Astra and Claude Fable 5.1" is limited to Periodic's self-built XRD evaluation and does not indicate that Neon possesses superior general reasoning, programming, or language capabilities, nor should it be used to establish an overall model ranking.

Moreover, a 55.3% success rate means that nearly half of the most challenging samples were still not successfully resolved. Neon is better suited as a tool to augment scientists' analytical capabilities rather than as an unreviewed fully automated materials expert.

Periodic's core logic is that high-quality scientific data available on the public internet is ultimately limited, whereas physical laboratories can continuously generate new data.

Traditional papers tend to favor the publication of successful results. Information about which materials were not synthesized, which temperatures and ratios failed, or what anomalies occurred with instruments is often excluded from papers, yet it is highly valuable for guiding the next round of experiments. An autonomous lab can retain all results—successful, failed, and uncertain.

The company breaks down material discovery into three cyclical steps:

First, predict what material should be synthesized; then predict how to synthesize it; finally, analyze what was actually produced in the experiment. The analysis results are fed back to the model to refine the next hypothesis.

Neon is currently focusing on improving step three. The next phase for Periodic aims to enable the model to directly plan research projects, write synthesis procedures, and select the next most informative experiment.

This is a key difference from mathematics and code reinforcement learning, where programs can run tests in seconds and mathematical answers can be verified by formal systems; physical experiments require equipment, materials, and time—sometimes lasting days—with results that may be ambiguous. Companies cannot instantly launch tens of thousands of real experiments like generating code, so they must make full use of accumulated data during experiment wait times to train and improve their models.

Periodic stated that Neon’s last training run peaked at 1,300 Nvidia H200 GPUs, covering both mid-term training and reinforcement learning phases. The company contrasted this with the over 100,000 Blackwell GPUs reportedly used by GPT-6 Astra, to illustrate how specialized data can reduce the computational power needed to train general scientific capabilities.

But this is not a like-for-like comparison.

Neon inherits the trillion-parameter weights of Kimi K2.6, along with all the data and computational resources previously invested in its pre-training. The 1,300 H200 GPUs represent only the peak scale during Periodic’s continued training of the model, not the total cost required to acquire this trillion-parameter capability. The company has not disclosed the full training duration, cumulative GPU hours, energy consumption, or total expenses.

More valuable are its engineering details.Infrastructure articlestates that the company modified Megatron, SGLang, Miles, and Ray to achieve 4.1 times the training throughput of its Megatron baseline and increased the decoding speed of trillion-parameter models from 10 tokens per request per second to 25.

An XRD Agent inference may last over an hour, during which scientific programs must be executed repeatedly. Early systems caused the model-generated code to request 80 GB of memory, leading to total training task failures. The team subsequently developed pbox, an isolated environment that runs model-generated code in a sandbox and utilizes idle CPU resources on GPU servers to handle scientific tools.

(The original infrastructure includes over an hour of Agent operation traces, an asynchronous training and inference architecture, and pbox sandbox performance comparison charts.)

Periodic also compared the two agent environments using the same Claude Opus 5.

One is Claude Code combined with an open-source crystal database and standard XRD software; the other is Periodic’s own scientific work environment, which includes experimental context, an internal materials database, simulation results, and custom analysis tools.

Under identical models and similar per-analysis costs, the success rate in the Periodic environment is 3.8 times higher than the former. This indicates that the bottleneck in XRD tasks is not only the model weights, but also whether the model can access correct experimental records, invoke appropriate tools, and retain prior research findings.

Therefore, Neon’s competitive advantage may be difficult to fully replicate by simply copying a model. The more valuable assets are the laboratory data, databases, instrument interfaces, and analytical processes behind the model.

This also suggests that scientific AI may create a competitive landscape distinct from general-purpose chat models: the larger the lab, the more proprietary data it generates; the better the model, the more experiments it can guide; and new experiments further expand the data advantage.

Complex XRD analysis typically lacks inexpensive, unique, and automatically verifiable standard answers. A well-fitted peak shape does not guarantee results that conform to chemical principles. Therefore, Periodic had three PhDs specializing in materials science independently annotate thousands of patterns, then trained a judging system composed of Claude Opus 5 and GPT‑5.6 Sol to score the model’s outputs.

The pairwise agreement rate among three human experts was 77.2%; the agreement rate between the LLM judge and a single human expert was 74.6%; and the agreement rate with the expert consensus was 84%.

These results indicate that AI judges can approximate expert judgment but are not equivalent to objective truth. Human evaluators naturally disagree, and裁判 models may favor certain response styles or overlook errors that all annotators failed to identify.

More importantly, FrontierXRD’s model, working environment, samples, and scoring methodology were all designed by Periodic. Although the company has published numerous methods and statistics, it has not released the full evaluation set, training data, or Neon weights that would allow external parties to reproduce the results, and the findings have not undergone peer review.

Periodic explicitly states that Neon was further trained from the open-weight model Kimi K2.6. As a result, some social media channels have referred to Neon as an "open-weight scientific model."

However, as of the time of retrieval, the official release does not provide Neon’s model card, license, or weights download link, nor does it offer code or data to fully reproduce the intermediate training and reinforcement learning. A more accurate statement is: Neon is built upon open-weight models, rather than having its trained Neon weights officially released to the public.

This distinction is important. External researchers currently cannot independently verify the 55.3% result, nor can they determine how much of the improvement comes from experimental data, reinforcement learning, specialized tools, reasoning budget, or evaluation environments.

Past scientific large models primarily learned knowledge from already published papers, patents, and database records. Periodic aims to take the next step: enabling the model to observe experiments completed today that have never appeared on the internet, and using these results to guide tomorrow’s experiments.

It is not yet a complete "AI scientist." Humans still need to define research goals, build laboratories, maintain equipment, design rewards, review XRD conclusions, and decide which materials are worth further investment. Neon has currently demonstrated that one key step can be significantly automated: as the number of experiments increases, analysis of results no longer requires a proportional increase in human experts.

If this closed loop works, the future competitive advantage of scientific models may no longer depend solely on who has more public text and GPUs, but on who can reliably run real experiments, document failed results, and turn the natural world into a continuously updated training environment.