Sage S2.1 Flash: Shanghai AI Lab's 321B Open-Weight Science Model
Sage S2.1 Flash is a 321-billion-parameter scientific multimodal model from Shanghai AI Laboratory, posted to Hugging Face under the MIT license around 15 November 2026. It is a mixture-of-experts model with about 18B active parameters, built on the GLM-5.3-Flash architecture. Its BF16 weights take about 645 GB, and the reference setup uses eight H200 GPUs.
On this page
A quiet release from Shanghai AI Laboratory
Shanghai AI Laboratory's InternLM team has published Sage S2.1 Flash on Hugging Face, describing it as a scientific multimodal model for research tasks that take several rounds of analysis and checking. Both the code and the weights are released under the MIT license, one of the most permissive licenses in common use.
There was no launch post. OrcaRouter, a model gateway that tracked the upload, reports that the main repository appeared on 15 November at 02:34 UTC, with an FP8 version created two days earlier, and that it could find no blog post or press release. AI Weekly, picking the model up on 16 November, likewise notes there is no paper, no description of the training data, and no stated release date. The Hugging Face collection groups the Flash model with a much larger 753B Sage S2.1 model and FP8 copies of both.
What the model is built for
The model card pitches Sage S2.1 Flash as a research agent rather than a plain chatbot. It combines understanding of scientific data with search, programming and tool use, and it is meant to run specialist software, check its intermediate results, and adjust later steps.
A companion component, Sage SciPerception, adds encoders and projectors for more than ten kinds of scientific input, including molecular graphs, crystal structures, protein sequences and 3D protein structures. The card lists tasks such as property prediction, classification, site localization and structure inference. Beyond science, it also names coding, office work, security and computer use as target uses.
The benchmark claims, with caveats
The card compares Sage S2.1 Flash with GLM-5.3-Flash, DeepSeek-V4.1-Flash, Qwen3.8-Max, GPT-5.6 Sol and Claude Opus 5. Its headline claim is first place on MLE-bench Lite with 90.3, just ahead of DeepSeek-V4.1-Flash at 89.2, plus top scores on OfficeQA Pro (72.2), Harvey LAB (25.0) and SpreadsheetBench 2 (37.1).
The same table shows where it trails. Claude Opus 5 leads on Terminal-Bench 2.1, SWE-bench Pro and OSWorld-Verified, and Qwen3.8-Max leads on PaperBench. All of these numbers are self-reported, and the card does not describe the evaluation setup. OrcaRouter notes that none had been independently reproduced at the time of writing, so treat them as claims until outside tests appear.
Can you run it locally?
Not on a typical home machine. The deployment guide describes a 320B-total, 18B-active mixture-of-experts model that reuses the GLM-5.3-Flash architecture, with BF16 weights of about 645 GB. Its reference setups use eight Nvidia H200 GPUs with vLLM or SGLang, and on Hopper GPUs the key-value cache must stay in BF16 because FP8 caching is not supported for this model.
OrcaRouter reports, from the model's config, a context ceiling of about one million tokens and a vision tower for images. Serving also needs the trust-remote-code flag, a patched vLLM build for some image and tool-call cases, and an explicit chat template for SGLang.
The low active-parameter count is the interesting part for local AI. Only about 18B parameters work on each token, so once community quantizations appear, large-memory workstations and multi-GPU rigs may run it at usable speeds. Until then, this is a model for labs and hosted providers. For anyone who wants a permissively licensed science agent they can inspect, fine-tune and run on their own hardware, it is a notable release; independent benchmarks and quantized builds are the next things to watch.