Shanghai AI Lab Releases a 320B Scientific Model Under the MIT License
Shanghai AI Laboratory released Sage S2.1 Flash, a multimodal mixture-of-experts model aimed at scientific work. It has 320 billion parameters and activates 18 billion per token. The BF16 checkpoint is about 645 GB across 120 shards. An FP8 release is approximately 330 GB.
What the benchmark reports
The source article lists a score of 90.3 on MLE-bench Lite, versus 89.2 for the next model in the table it cites. That finding concerns the benchmark and setup used in the report. It does not establish equal accuracy for every scientific domain or remove the need to check experimental outputs.
Why deployment size matters
Sparse activation reduces the compute needed for a token, while all relevant weights still must be available to the serving system. Plan for storage, accelerator memory, interconnects, and runtime overhead. Check the deployment guide for the exact checkpoint and inference configuration.
Which release this is
S2.1 Flash is distinct from the earlier Sage S2.1. The weights appeared on November 15, with supporting model and deployment material added November 16. Review those files and try a sample of your own tasks before committing to a large download.