The Shanghai Artificial Intelligence Laboratory published Sage S2.1 Flash on Sunday (Nov 15), a multimodal model built for scientific work, with 321 billion parameters and code and weights released under the MIT license. The model is already available on Hugging Face and ModelScope, and the technical report was submitted to arXiv on Monday (Nov 16).
A quiet release
Unlike this week's big model launches, Sage S2.1 Flash arrived with no announcement post, no briefing and no press release: the repository simply appeared on Hugging Face on November 15, and the project microsite only showed up the next day. It was AI-focused press that spotted the publication and detailed what the lab had put out.
What the model does
Sage S2.1 Flash was designed as a research agent, not an ordinary chatbot. It combines scientific data understanding with information retrieval, programming and tool use, and it can operate scientific software such as Blender, 3D Slicer and Fiji, checking intermediate results against visual feedback before deciding the next step.
At its core is Sage SciPerception, a unified representation layer that accepts more than ten types of scientific input — molecular graphs, crystal structures, protein sequences and 3D structures, DNA, RNA, spectra, MS/MS, XRD, ECG, EEG and light curves — and projects them into the same representation space as text, so the model can reason about them alongside natural-language instructions. Context reaches about 1 million tokens.
Numbers and benchmarks — with a caveat
According to the lab's own evaluation, the model placed first on 4 of 15 benchmarks and second on 7. The headline result is MLE-bench Lite, with 90.3 points, ahead of DeepSeek-V4.1-Flash-1017 (89.2). The model also leads OfficeQA Pro (72.2), Harvey LAB (25.0) and SpreadsheetBench 2 (37.1), as well as PostTrainBench (39.2). The comparisons include GLM-5.3-Flash, Qwen3.8-Max, GPT-5.6 Sol and Claude Opus 5.
The caveat matters: all of these numbers are self-reported by InternLM, with no independent reproduction so far and no evaluation harness files published alongside the model. They should be read as a developer claim, not as a third-party verified result. An independent analysis of the configuration files also indicates the architecture reuses the GLM-5.3-Flash base.
MIT license and where to download
Code and weights are under the MIT license, one of the most permissive, with BF16 and FP8 versions on Hugging Face and ModelScope. In practice, though, the model is heavy: 645 GB of weights in BF16, with 18 billion active parameters per token, and the reference setup uses eight Nvidia H200 GPUs with vLLM or SGLang. It is a model for labs and hosting providers, not for a home machine. The same Hugging Face collection also includes Sage S2.1, a larger version with 753 billion parameters.
Why it matters
With open weights under MIT and native scientific data support, Sage S2.1 Flash enters a race so far dominated by closed models and specialized tools: an AI assistant that reads molecules, proteins and spectra as direct input, rather than relying on text alone. If the numbers hold up in independent testing, it could become a relevant option for labs that need to run sensitive research inside their own security perimeter.
Read also: OpenFold3 arrives on SageMaker JumpStart, bringing biomolecular structure prediction to the AWS cloud.
Fontes
- Hugging Face — Sage S2.1 Flash (model card oficial)
- Sage S2.1 Flash — site oficial do projeto
- Relatório técnico — arXiv 2611.21072
- CuriousLM — análise do lançamento
- AI Weekly — InternLM's 321B Sage S2.1 Flash leads MLE-bench Lite at 90.3
- TechFastForward — InternLM Sage 2.1 tops MLE-Bench at 90.3 with MIT license
Image: diagram of the Sage SciPerception architecture, from the model's official repository on Hugging Face (InternLM / Shanghai Artificial Intelligence Laboratory), licensed under MIT.
Transparency: this article was produced with the assistance of artificial intelligence, based on cross-checking public posts and the internet sources cited above. We recommend consulting the original sources to confirm the details.
By Muse