--- id: "2026-11-15-shanghai-ai-lab-sage-s2-1-flash-open-weights" url: "https://postcutoff.com/e/2026-11-15-shanghai-ai-lab-sage-s2-1-flash-open-weights/" as_of: "2026-11-22T13:32:00+01:00" date: "2026-11-15" date_precision: day category: open-source importance: 2 confidence: medium status: [Partly confirmed] sources: 3 editor: human editor human_review: null version: "2026-11-18" --- As of: 2026-11-22 13:32 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-11-15-shanghai-ai-lab-sage-s2-1-flash-open-weights/ # Sage S2.1-Flash open weights Full title: Shanghai AI Lab quietly posts Sage S2.1-Flash, a ~320B open-weight scientific multimodal MoE model (MIT licence) On 15 November 2026 the InternLM organisation of the Shanghai AI Laboratory uploaded Sage S2.1-Flash to Hugging Face under the MIT licence, without a press release. OrcaRouter describes it as a 320B-total, 18B-active MoE in bfloat16 (~645 GB); AI Weekly reports it handles 10+ scientific input types (molecular graphs, crystal structures, protein sequences and structures) and leads MLE-bench Lite at 90.3 ahead of DeepSeek V4.1 Flash (89.2). The benchmark claims are single-source and unverified here. ## Key facts - Hugging Face repo internlm/Sage-S2.1-Flash created 2026-11-15 02:34 UTC, last modified Nov 16; model card body empty at the time OrcaRouter checked; architecture tag 'glm5_next' (HF API, read Nov 18: 678 downloads, 125 likes) - Reported size: 320B total / 18B active (OrcaRouter) or 321B (AI Weekly); BF16 weights plus an FP8 variant (AI Weekly) - Claimed: MLE-bench Lite 90.3, first, vs DeepSeek-V4.1-Flash-1017 89.2 (AI Weekly; results in a separate GitHub repo per OrcaRouter) ## What happened A large Chinese open-weight science model appeared without announcement; found through a web search on 18 Nov, not through the sweep (no lab feed or Techmeme story). ## Your AI and this story - GPT-6.2 Astra (training cutoff May 2026): 168 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 138 days after its cutoff - Gemini 4 Hydrogen (training cutoff March 2026): 229 days after its cutoff - Grok 4.7 (training cutoff May 2026): 168 days after its cutoff ## Sources 1. [Hugging Face: internlm/Sage-S2.1-Flash](https://huggingface.co/internlm/Sage-S2.1-Flash) (huggingface.co, code) 2. [AI Weekly: InternLM's 321B Sage S2.1 Flash leads MLE-bench Lite at 90.3](https://aiweekly.co/alerts/internlms-321b-sage-s21-flash-leads-mle-bench-lite-at-903) (aiweekly.co, press) 3. [OrcaRouter: Sage S2.1 Flash's quiet release](https://www.orcarouter.ai/blog/sage-s2-1-flash-quiet-release) (orcarouter.ai, discussion) ## Changes - 2026-11-18 (filed): Created. Benchmark claims single-source