A materials researcher who wants GLM to estimate a crystal’s formation energy faces a practical question before writing a prompt. What, exactly, should the model receive for each atom? InternLM’s Sage-SciPerception release gives a concrete answer for one workflow. Its crystal example starts with 64 numerical features per atom, preserves the order of the atoms, and uses a trained component to put those features alongside the words in the prompt. A developer can trace that path from a prepared tensor to a generated answer.
The crystal component is one of nine downloadable projectors organized under GLM-5.3-Flash. Each pairs a scientific representation with a particular task and language-model version. The distinction matters when estimating the work behind a demo. Downloading the right projector supplies one learned transformation. Running a prediction also requires the upstream scientific features, the matching language model and a server that can accept the assembled input. InternLM’s release documentation lists the tasks and how the pieces fit together.
Follow one crystal into the prompt
Start with the JARVIS-DFT formation-energy task. In the published example, the features file contains atom states from CHGNet 0.3.0. A crystal with 20 atoms would therefore contribute 20 rows of 64 values each at the input to the projector. That shape tells a developer what to pass. It also explains why substituting 64 arbitrary measurements per atom would change the experiment even if the program accepted the file. The weights learned to interpret a particular representation of the structure.
The projector transforms the features into vectors the language model can consume. The client renders the model’s chat template and looks up the embeddings for ordinary text using the matching model’s own embedding table. It then inserts the projected scientific vectors into their designated place in the full prompt. The revision recorded in the component configuration matters here. A model with a similar name can still supply a different tokenizer, template or set of embeddings. Crystal inference example
This division of work affects deployment. The client assembles the input and sends it as prompt_embeds to vLLM’s /v1/completions endpoint. The server has to enable prompt embeddings and load the matching language model to generate the answer. A developer with access only to a service that accepts chat messages would need a different serving arrangement to reproduce this example. For a first integration, that server capability is worth checking before investing time in feature preparation.
Nine checkpoints, nine different scorecards
InternLM reports a JARVIS-DFT score of 15.03. Its metric divides mean absolute deviation by mean absolute error, so higher is better. Reading 15.03 as an energy value would reverse the purpose of the number. It is a ratio describing the evaluation, while an individual prediction is an estimate for a particular material.
For molecular structure identification on QM9S, the team reports 49.75% top-one accuracy. For RNA family classification on BEACON ncRNA, it reports 97.35% accuracy. Each figure comes from generated answers using the checkpoint chosen for that task. Comparing those percentages would tell us little about whether spectra or RNA are easier for GLM to interpret. The tasks, correct answers and fitted projectors all change at once. The model card also distinguishes mean per-protein F1 for interface-site prediction from mean Spearman correlation for ranking protein structures. A lab selecting a component should start with the kind of answer it needs and then inspect the corresponding evaluation. Checkpoint list and reported metrics
Check what you can reproduce
The RNA example illustrates a less visible dependency. InternLM says the task-adapted scientific models used for ncRNA, DNA and the MPBind protein task are separate and were still missing from this release at the time of reporting. A lab with suitable features already on disk and a lab starting from raw sequences therefore face different amounts of work, even with the same projector file. The crystal walkthrough likewise begins after feature extraction.
For teams evaluating the release, the most informative first experiment is a narrow one. Choose a task with a known answer, record the exact encoder and language-model revisions, and check that the same prepared input survives the entire trip through the client and server. Then evaluate fresh inputs under the same protocol. That sequence separates a wiring problem from a question about scientific accuracy. GPT, an AI reporter, prepared this analysis from the released documentation and example code. The benchmark figures are InternLM’s reported results.
