THE ESSENTIALS IN 30 SECONDS

  • Sage S2.1 Flash has 321 billion parameters and is released under the MIT licence for both code and weights, with an FP8 variant for lighter deployment.
  • Through a companion encoder, it accepts more than ten scientific input types, including molecular graphs, crystal structures and protein structures.
  • The model card reports a 90.3 score on MLE-bench Lite, but gives no paper, training-data description, context length or active-parameter count.

Shanghai Artificial Intelligence Laboratory, through its InternLM organisation, has put a new large model for scientific research on Hugging Face. Called Sage S2.1 Flash, it has 321 billion parameters and is licensed under MIT, which covers both the code and the weights. It is also listed on ModelScope.

A quiet release

There was no launch announcement. According to OrcaRouter, which tracked the repositories, the main BF16 repository was uploaded on November 15 at 02:34 UTC with roughly 645 GB of weights, while an FP8 version had been created about 35 hours earlier. A GitHub repository hosting a demo page and the benchmark table followed on November 16. OrcaRouter found no mention of the model on the lab's homepage or in a press release at that point.

What the model card describes

The lab presents Sage S2.1 Flash as a scientific multimodal model. A companion encoder, Sage SciPerception, lets it take in more than ten kinds of scientific data alongside text, such as molecular graphs, crystal structures, protein sequences and 3D protein structures. AI Weekly describes the intended use as a research assistant that combines data understanding, retrieval, coding and tool use for multi-step analysis.

All performance figures are the lab's own. The model card states that, across 15 benchmarks and six models, Sage S2.1 Flash ranks first on four and second on seven. It reports 90.3 on MLE-bench Lite, ahead of 89.2 for DeepSeek-V4.1-Flash-1017, and first places on OfficeQA Pro (72.2), SpreadsheetBench 2 (37.1) and Harvey LAB (25.0). It trails Claude Opus 5 on Terminal-Bench 2.1 (87.9 against 90.2) and OSWorld-Verified (85.2 against 86.4), and Qwen3.8-Max on BabyVision (75.8 against 83.8).

What is missing

Several basic details are absent. The card does not name the base model, describe the architecture, give a context length or say how many parameters are active per token. There is no accompanying paper and no description of the training data. The card is also not fully consistent: its comparison table covers many more than 15 benchmarks and highlights five first-place results rather than four. The scores are self-reported and should be read as such until outside evaluations appear, and running the full model requires hardware able to hold several hundred gigabytes of weights.