HomeNews

InternLM Sage S2.1-Flash: 321B Scientific Model Tops MLE

Kyle Morrison
Kyle Morrison
•
Nov 16, 2026 6:43 PM
•
0
 min read
Select Emergent as your Preferred news source
InternLM Sage S2.1-Flash: 321B Scientific Model Tops MLE

💡 TL;DR

  • Shanghai AI Lab released Sage S2.1-Flash, a 321B-parameter scientific multimodal model under MIT license on Hugging Face.
  • The model scored 90.3 on MLE-bench Lite, claiming first place ahead of DeepSeek-V4.1-Flash and topping four of fifteen benchmarks.
  • Sage SciPerception encoder supports molecular graphs, crystal structures, protein sequences, and over ten scientific input types.

Shanghai Artificial Intelligence Laboratory has released Sage S2.1-Flash, a 321-billion-parameter scientific multimodal model, on Hugging Face under an MIT license. Officially launched on November 16, 2026, the system ships with both BF16 and FP8 weight variants, targeting researchers who need to process molecular structures, protein data, and other scientific inputs alongside natural language. The model card reports a 90.3 score on MLE-bench Lite, placing it ahead of DeepSeek-V4.1-Flash-1017 at 89.2, and claims first-place rankings on four of fifteen evaluated benchmarks.

Architecture and Scientific Perception Stack

Sage S2.1-Flash centers on a representation system called Sage SciPerception, which encodes more than ten types of scientific input into a shared latent space. The model card lists molecular graphs, crystal structures, and protein sequences and 3D structures among the supported formats. Each input type passes through a dedicated encoder, then projects into the common embedding space where the language model can reason over scientific and textual tokens together. Downstream capabilities include prediction, classification, site localization, quantitative analysis, and structure inference.

InternLM frames the release as a research assistant rather than a conversational agent. According to the card, the model is designed for tasks requiring multiple rounds of analysis and verification, combining scientific data understanding with information retrieval, programming, and tool use. It can invoke specialized software, check intermediate outputs against visual feedback, and adjust subsequent steps dynamically.

Benchmark Performance

The model card reports first-place finishes on MLE-bench Lite, OfficeQA Pro, Harvey LAB, and SpreadsheetBench 2, with scores of 90.3, 72.2, 25.0, and 37.1 respectively. Additional benchmarks include Terminal-Bench 2.1 at 87.9 and DeepSWE at 67.9. Across fifteen benchmarks and six comparison models, Sage S2.1-Flash placed first on four and second on seven. The numbers come from InternLM's internal evaluation and have not been independently reproduced.

No accompanying paper or training-data description accompanies the release. The card does not specify dataset composition, training duration, or compute budget. The MIT license permits commercial use and redistribution without restriction.

Availability and Deployment Options

Both BF16 and FP8 weight variants are available on Hugging Face as of today. The FP8 version reduces memory footprint for teams running inference on consumer hardware or limited GPU clusters. InternLM has not announced API endpoints or hosted inference options, so practitioners must host the model locally or on private cloud infrastructure.

The release follows a steady stream of large open-source models this quarter, including several scientific and tool-use-focused systems. Sage S2.1-Flash's combination of parameter scale, permissive licensing, and scientific domain support positions it as a candidate for academic and industry research pipelines that require multimodal reasoning over technical data.

What This Means

Sage S2.1-Flash extends the frontier of open scientific models by packaging domain-specific encoders with a large language backbone under a permissive license. Its top performance on MLE-bench Lite and competitive scores on reasoning and spreadsheet tasks suggest utility beyond narrow scientific workflows. Researchers can now process molecular graphs, protein structures, and natural language in a single inference pass without proprietary APIs. The FP8 weights lower the barrier for teams without enterprise GPU budgets, and the MIT license removes deployment friction for commercial projects. Expect rapid adoption in computational biology, materials science, and hybrid research environments where proprietary models remain off-limits.

About the writer

Kyle Morrison leads Community at Emergent, where he brings together founders, operators, and SMB owners who are building real software with AI agents. He focuses on helping builders go from their first app to production use.

Start Building
on Emergent today
Try Emergent