A high-throughput facility in Menlo Park developed a specialized artificial intelligence that outperforms general frontier systems in the analysis of semiconductor materials. Periodic Labs calls the tool Neon, a 1 trillion parameter open-source LLM trained via mid-training and reinforcement learning. The model used 1,300 H200 GPUs and months of wet lab data to surpass GPT-6 Astra and Fable on benchmarks for superconductors and magnets.
The laboratory infrastructure reached 4.1x training throughput over a Megatron baseline and 2.5x faster decoding with cluster utilization above 95%. This architecture creates a cycle where AI manages lab operations and learns from the resulting data to inform future experiments. Periodic Labs contributes its technical improvements to open-source projects including Megatron-LM, SGLang and Miles.
Key sources
- SOURCE@liamfedus“Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark”x.com
- SUPPORT@liamfedus“train specialized scientific models with strong performance on relatively modest compute compared to frontier systems”x.com
- SUPPORT@liamfedus“4.1× training throughput vs. our Megatron baseline 2.5× faster decoding 95%+ cluster utilization”x.com
- SUPPORT@khoomeik“We midtrain + RL’d a trillion param LLM to analyze experimental data from our superconductor lab”x.com
- SUPPORT@_jasonwei“wet lab data enables a specialized model to beat gpt-6 astra at a task at the frontier of science!”x.com
- SOURCEmarketbrief.now