Sage S2.1 Flash: Shanghai AI Lab's 320B science model that reads molecules and spectra
Shanghai AI Lab quietly shipped a 320B/18B-active MoE science model under MIT license that takes molecules, protein structures, spectra and EEG as inputs alongside text, topping MLE-bench Lite at 90.3.
Why it matters
The first open-weight model built science-native rather than chat-first: a unified SciPerception encoder maps 10+ scientific input types into the same embedding space as text, and its self-reported MLE-bench Lite 90.3 leads DeepSeek-V4.1-Flash - though no independent harness has reproduced the numbers yet.
What you could build with it
A computational biology team could build a literature-to-lab assistant on Sage S2.1 Flash's SciPerception encoders, feeding it a molecule or protein structure plus the relevant papers and letting it propose and check candidate compounds, then sell the screening service to small biotech firms that cannot staff a full cheminformatics team.
Does it hold up?
Too early to judge - weights are live under MIT and the deployment guide names an 8xH200 reference setup, but the model card is a one-line licence, no inference provider serves it yet, and all benchmark figures are self-reported with no independent reproduction.
Built with Sage S2.1 Flash
- InternLM Sage 2.1 Tops MLE-Bench at 90.3 With MIT Licensetechfastforward · Report on the quiet Nov 15 weights drop: 645 GB across 120 shards, MIT license, no blog post or press release from the lab, and the 90.3 MLE-bench Lite claim ahead of DeepSeek-V4.1-Flash-1017 at 89.2.
- Sage-S2.1-Flash vs DeepSeek V4.1 Flash: one shared tableorcarouter · Independent read of the release: 320B/18B MoE on the GLM-5.3-Flash architecture, 1M-token ceiling, 8xH200 reference deployment, and the caveat that all 35 benchmark rows are self-reported in a GitHub repo the HF card never links to.
- internlm/Sage-S2.1-Flash on Hugging Facevibingtalk · Community mirror of the HF release: science multimodal model for multi-round research tasks, first on 4 of 15 benchmarks and second on 7, with the full evaluation table.
- InternLM/Sage-SciPerceptiongithub · 13 ★ · Official scientific perception toolkit: trained projectors that connect crystals, spectra, biological sequences and biosignals to the language model, with the inference guide the HF deployment docs point to.
Learn more
First spotted on hf: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Google AI Edge Gallery: offline speech recognition and synthesis on-deviceGoogle added fully offline speech recognition and text-to-speech to AI Edge Gallery, with ML Kit APIs on Android and…models · JEV 0.73Soup: fine-tune an 8B LLM on a 4 GB laptop GPU from one YAML fileAn open-source CLI that fine-tunes LLMs from a single YAML file, with layer streaming that trains an 8B model on a 4…models · JEV 0.73Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.