Back home

书生 S2.1 Flash / Sage-S2.1-Flash

x+web 0.92
Visit source

Intro

Sage-S2.1-Flash (Chinese official name 书生 S2.1 Flash) is a scientific multimodal agent model formally released on 16 November 2026 by Shanghai Artificial Intelligence Laboratory’s InternLM / 书生 line. The official page at internlm.github.io/Sage-S2.1-Flash opens with “今天,我们正式发布 书生 S2.1 Flash” (“Today, we formally release 书生 S2.1 Flash”). It describes a frontier scientific multimodal agent model for demanding research: a 320-billion-parameter MoE that combines scientific multimodal understanding with general agent capabilities, for scientific deep research, scientific programming, and operation of specialized research software.

The lab says the model unifies more than ten kinds of scientific input — including molecular graphs, crystal structures, protein sequences and 3D structures, DNA/RNA sequences, spectra, mass spectra, X-ray diffraction, ECG, EEG, and light curves — under an architecture named Sage SciPerception. It also says the model was trained in large-scale verifiable environments to strengthen search, programming, tool use, and multi-step execution, and can inspect software state and intermediate results via visual feedback before revising its next steps. Official demos include recovering a single CT series from mixed slices and reconstructing bone in 3D Slicer, 3D modeling in Blender / CAD, nucleus segmentation in Fiji, and interactive literature-backed scientific analysis.

The official evaluation table covers 30-plus benchmarks. Among them it lists 90.3 on MLE-bench Lite, 72.2 on OfficeQA Pro, 25.0 on Harvey LAB, and 37.1 on SpreadsheetBench 2, first on those four among the comparison set shown (GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Qwen3.8-Max, GPT-5.6 Sol, and Claude Opus 5). Those figures are the lab’s own evaluations; this page does not independently reproduce them.

Jiqizhixin reported at 19:37 Beijing time the same day that the model was jointly released by Shanghai AI Laboratory with Fudan University, the Institute of Software of the Chinese Academy of Sciences, Renmin University of China, the Institute of Automation of the Chinese Academy of Sciences, East China Normal University, Harbin Institute of Technology, and Peking University, and linked a demo, Hugging Face, ModelScope, and a technical report. The same article also described a limited-time API promotion against GLM 5.3 Flash: 1% of the comparison model’s price through 9 December, then 10% from 10–31 December, plus a new-user credit grant — confirm current terms on the official demo page. Weights are released under the MIT license. @ModelScope2022 posted at 2026-11-16T11:28:20Z (19:28:20 Beijing time) that @sageai_official had open-released Sage-S2.1-Flash, describing it as a 320B-A18B multimodal MoE.

The previous 16:32 scan already saw the GitHub repository InternLM/Sage-S2.1-Flash (created at 2026-11-16T03:34:03Z, about 9 stars then) but found no formal announcement, so it was not stored. This slot has the official release page and the Jiqizhixin report, so the name is filed separately.

Key points:

Advertisement

Explain

Why it matters

This is a new model name Shanghai AI Laboratory confirmed on its official release page, not a routine update to the existing Sage-S2.1 or earlier Sage terms. The weight repositories already appeared on GitHub and Hugging Face earlier on 16 November; the previous 16:32 scan did not store the name because no formal “released today” announcement had been found yet. In this 20:32 window the official page (“Today, we formally release 书生 S2.1 Flash”), the 19:37 Jiqizhixin report, and a ModelScope post attributing an open release to @sageai_official are available, so the name is filed separately.

This page only restates the official release page, repository materials, the ModelScope introduction post, and the same-day Jiqizhixin report of the official figures. Scores such as 90.3 on MLE-bench Lite are company statements from the official table; this page does not independently reproduce them.

Limits

The limited-time API promotion (1% / 10% of the GLM 5.3 Flash comparison price, plus new-user credits) should be rechecked on the official demo page. Do not treat this page as a guarantee of quality, scientific-software skill, or leaderboard rank. Do not describe this model as the existing terms Sage-A1.5 or Sage2, or as the earlier Sage-S2.1 (non-Flash).

Usage

  1. Read the official release page at https://internlm.github.io/Sage-S2.1-Flash/ and cross-check the demo at https://discovery.intern-ai.org.cn/token-plan/home ; weights and docs are at https://huggingface.co/internlm/Sage-S2.1-Flash , https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Sage-S2.1-Flash , and https://github.com/InternLM/Sage-S2.1-Flash ; a same-day report is at https://finance.sina.com.cn/roll/2026-11-16/doc-iniyxufc2704181.shtml ;
  2. Follow the official instructions to try the hosted demo, or deploy the MIT-licensed weights yourself; confirm any limited-time pricing or new-user credits on the official demo page ;
  3. When citing the 320B MoE / 320B-A18B figures, the ten-plus scientific input types, Sage SciPerception, or official scores such as 90.3 on MLE-bench Lite, attribute them to the 16 November 2026 official page, repository materials, the ModelScope post, or reporting of those official figures, and recheck the upstream source ;
  4. Do not describe 书生 S2.1 Flash / Sage-S2.1-Flash as the existing terms Sage-A1.5 or Sage2, or as the earlier Sage-S2.1 (non-Flash); this page only restates official materials and reporting — recheck the upstream sources before citing them externally.