AI for Science／訓練基礎設施
Periodic Neon Trains a Trillion-Parameter XRD Agent on Laboratory Data and Reveals the System Bottlenecks of Long-Trajectory RL
Periodic Labs used 1,300 H200 GPUs to conduct scientific mid-training and reinforcement learning on a trillion-parameter open model, enabling Neon to automatically analyze complex X-ray diffraction results. Its 55.3% success rate and comparisons with frontier models are all based on internal data, a proprietary toolchain, and model-based judging. At this stage, it is better viewed as a testable vertical AI recipe than as a demonstrated system for autonomous scientific discovery.
Periodic Labs has introduced Periodic Neon, starting from a trillion-parameter open-weight model and adding multimodal mid-training on papers, code, and its own experimental data, followed by long-context reinforcement learning using laboratory data. Its first deployment task is powder X-ray diffraction (XRD) analysis: the agent must combine diffraction peaks, synthesis conditions, historical experiments, crystal databases, and simulation tools to determine which crystalline phases are present in a sample and in what proportions, then feed the results back into the materials-discovery loop.
On 134 FrontierXRD samples that would each take human experts several hours to process, the company reports a 55.3% success rate for Neon—roughly 20 times the 2.7% achieved by the base Kimi K2.6 model—and says it outperformed GPT-6 Astra and Claude Fable 5.1 under its cost assumptions. Another 198 chemical systems excluded from both mid-training and RL were used for generalization testing. Training peaked at 1,300 H200 GPUs, but this was not the total cost of pretraining from scratch; it was the cost of additional training performed on top of existing trillion-parameter weights.
The most broadly applicable part of this release is actually its system design. An XRD rollout can last more than an hour, while a training step takes only a few minutes. The team therefore ran generation and updates asynchronously on separate GPU allocations, allowing a single trajectory to span multiple weight versions. Increasing KV cache capacity with WideEP amplified synchronization stalls across data-parallel ranks; the team switched to Delta router replay, which returns routing data only for newly generated tokens, and contributed the relevant changes to SGLang. Combined with prefill/decode disaggregation, the company says this increased the trillion-parameter model's throughput on the same batch from 10 tok/s to 25 tok/s per request. Long-sequence packing and cost-aware scheduling raised training throughput to 4.1 times its Megatron baseline.
The evaluation still warrants cautious interpretation. Every model used Periodic's proprietary harness, which includes internal databases and experimental context; the same model can achieve a success rate 3.8 times higher with this harness than with a general-purpose toolset. Success labels were also assigned by a judging panel composed of Opus 5 and GPT-5.6 Sol. Although its decisions agreed with expert consensus 84% of the time, this is still not objective ground truth. Neon's weights, the raw FrontierXRD data, and the complete training recipe have not yet been released. The next question is whether it can progress from analyzing existing results to reliably selecting new experiments—and whether external laboratories can reproduce its claimed gains in materials discovery.