Is Sage S2.1 Flash state of the art?Pending benchmarks.Announced Nov 16, 2026; no independent results yet.
Shanghai Artificial Intelligence Laboratory's scientific multimodal model (书生 S2.1 Flash). Official materials describe a 320B-A18B mixture-of-experts that uses the GLM-5.3-Flash architecture. Through Sage SciPerception it takes more than ten kinds of scientific input — including molecular graphs, crystal structures, protein sequences and 3D structures, DNA and RNA sequences, spectra, mass spectra, X-ray diffraction, electrocardiograms, electroencephalograms, and light curves — alongside natural-language instructions, and combines scientific data understanding with retrieval, programming, and tool use. Announced November 16, 2026 (official website and InternLM GitHub repository created 2026-11-16T03:34:03Z, 10:34 PM ET November 15; team posts later the same calendar day in ET). Open MIT-licensed weights, ungated on Hugging Face at internlm/Sage-S2.1-Flash and on ModelScope, with an FP8 build. No documented public self-serve API model id was confirmed.
The details so far
- Lab
- Intern Large Models
- Announced
- Nov 16, 2026
- API
- Not yet
- Open weights
- Published · Ungated internlm/Sage-S2.1-Flash (BF16) and internlm/Sage-S2.1-Flash-FP8 on Hugging Face; mirrored on ModelScope under Shanghai_AI_Laboratory
- Parameters
- 320B-A18B MoE (official deployment guide); official website says 320-billion-parameter MoE. Hugging Face safetensors metadata totals 321,323,031,390 parameters
- Developer
- Shanghai Artificial Intelligence Laboratory (official model card and website)
- Architecture
- Official deployment guide: 320B-A18B multimodal MoE in bfloat16 (about 645 GB), using the GLM-5.3-Flash architecture. Official config.json: Glm5NextForConditionalGeneration / glm5_next, 45 layers, 288 routed experts plus 1 shared, 8 experts per token
- Context
- 1,048,576 tokens (max_position_embeddings in official config.json)
- Input
- Text and images, plus scientific inputs via Sage SciPerception: molecular graphs, crystal structures, protein sequences and 3D structures, DNA and RNA sequences, spectra, mass spectra, X-ray diffraction, electrocardiograms, electroencephalograms, and light curves (official website and model card)
- License
- MIT (official model card)
- Hosted access
- Official website links Try the API to https://discovery.intern-ai.org.cn/ (Intern InkStone). A documented public self-serve API model id was not confirmed without an account
- Related
- Separate from Sage-A1.5 (InternScience), Sage2 and Sage 1.3.0 (Levanto Labs), Atria Dawn, and GLM-5.3-Flash
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: self-reported official model-card scores of MLE-bench Lite 90.3, OfficeQA Pro 72.2, SpreadsheetBench 2 37.1, Terminal-Bench 2.1 87.9, DeepSWE 67.9, and SWE-bench Pro 60.6.
Sources
LABSage S2.1 Flash official websiteinternlm.github.io/Sage-S2.1-Flash/ · Nov 16, 2026REPOinternlm/Sage-S2.1-Flash on Hugging Facehuggingface.co/internlm/Sage-S2.1-Flash · Nov 16, 2026LABSage S2.1 Flash deployment guidehuggingface.co/internlm/Sage-S2.1-Flash/blob/main/deployment_guide.md · Nov 16, 2026REPOinternlm/Sage-S2.1-Flash-FP8 on Hugging Facehuggingface.co/internlm/Sage-S2.1-Flash-FP8REPOShanghai_AI_Laboratory/Sage-S2.1-Flash on ModelScopemodelscope.cn/models/Shanghai_AI_Laboratory/Sage-S2.1-FlashREPOInternLM/Sage-S2.1-Flash on GitHubgithub.com/InternLM/Sage-S2.1-Flash · Nov 16, 2026LABTeam post on X (quoted announcement)x.com/i/status/2122319410623029578 · Nov 16, 2026LABTeam post on Xx.com/wanghan_xu/status/2122378890689851575 · Nov 16, 2026Tracking continues until an independent board lists it.