Skip to content
, ,

Sage S2.1 Flash: Shanghai AI Lab’s 320B MIT Model That Reads Molecules and Spectra (Benchmarks, Hardware, Setup)

12–18 minutes
Woman in a science lab holding a tablet with a SAGE card, hook 320B open science AI

Shanghai AI Laboratory quietly put Sage S2.1 Flash on Hugging Face on 15 November 2026, and the technical report followed on arXiv on 16 November. It is a 320B-parameter mixture-of-experts model with about 18B active, released under the MIT licence.

What makes it different from the other big open models this month is science. Besides text and images, it can take molecules, crystal structures, protein and DNA sequences, spectra, ECG and EEG signals through small add-on “projectors”.

This guide covers the full specs, all 33 vendor benchmarks, the exact vLLM, SGLang and LMDeploy commands, what hardware it really needs, how the scientific inputs work, and how it compares with GLM-5.3-Flash, DeepSeek V4.1 Flash and Qwen3.8.

TL;DR

  • What: Sage S2.1 Flash, an open-weights scientific multimodal model from Shanghai AI Lab (InternLM team).
  • Size: 320B total (321B on the Hub), about 18B active, 288 experts with 8 active per token, built on the GLM-5.3-Flash architecture.
  • Licence: MIT for code and weights. Commercial use is allowed.
  • Downloads: internlm/Sage-S2.1-Flash (BF16, about 645 GB) and internlm/Sage-S2.1-Flash-FP8 (about 330 GB).
  • Hardware: 8x H200 for BF16, 4x H200 for FP8 (vendor examples). This is a server model, not a home-PC model.
  • Best scores (vendor-reported): MLE-bench Lite 90.3, OfficeQA Pro 72.2, SpreadsheetBench 2 37.1, Harvey LAB 25.0. It trails Claude Opus 5 on SWE-bench Pro (60.6 vs 76.5).
  • No hosted API price yet and no independent benchmark runs so far.

What Is Sage S2.1 Flash?

Sage S2.1 Flash comes from Shanghai Artificial Intelligence Laboratory, the state-backed lab behind the InternLM model family. The weights sit under the internlm organisation on Hugging Face and under Shanghai_AI_Laboratory on ModelScope.

The release was quiet. There was no launch blog post when the weights appeared, only the model cards, a deployment guide, a bilingual website (internlm.github.io/Sage-S2.1-Flash) and the arXiv paper “Sage S2.1 Flash Technical Report” (arXiv 2611.21072).

The paper describes it as a general-purpose model for two jobs:

  • Reasoning over scientific evidence: reading molecules, crystals, sequences and measurement signals, not just text about them.
  • Acting as an agent: tool use, web search, coding, running ML experiments and driving scientific software such as 3D Slicer, Blender and Fiji (shown in the vendor demo videos).

The agent training used “verifiable environments”. In plain words, the model got rewarded only when a task really finished, checked by execution results, the final state of the environment, or the files it produced.

Note: The paper’s abstract calls the model “Sage S1.5 Flash” in a few places, while the title and metadata say S2.1. It looks like a leftover from an earlier draft. The model you download is S2.1 Flash.

Full Specs

Item Sage S2.1 Flash
Maker Shanghai AI Laboratory (InternLM)
Weights published 15 Nov 2026 (Hugging Face); paper 16 Nov 2026
Parameters 320B total (Hub metadata: 321B), about 18B active (“320B-A18B”)
Architecture MoE on GLM-5.3-Flash design (glm5_next): 288 routed experts, 8 active per token, 1 shared expert
Layers 45 (34 linear-attention, 11 sparse-attention, per config analysis)
Inputs Text, images, plus 10+ scientific input types through projectors
Context Config allows up to 1,048,576 positions; no official context limit stated
Precision BF16 (about 645 GB, 120 shards) and FP8 E4M3 (about 330 GB)
Licence MIT (code and weights)
Serving engines vLLM 0.33.0 (with patch), SGLang 0.5.20+, LMDeploy (main branch)
Hosted API None with public pricing yet

The 18B active figure matters. Each token only runs through about 18B parameters, so it generates far faster than a 320B dense model would. But all 320B still have to sit in memory, which is why the hardware bill stays high.

The Scientific Inputs, Explained

This is the headline feature. Most chat models can only read a molecule as a text string such as a SMILES code. Sage uses a component called Sci-Perception that turns the raw data into something the language model can “see”.

It works in three steps:

  1. A specialist encoder reads the data. For example, crystal files go through CHGNet, protein sequences through ESM2, and DNA through DNABERT2 (per the projector configs). Spectra are cut into ordered patches of measurements.
  2. A small projector translates it. Each data type has a trained projector (only 4 to 28 MB each) that maps the encoder output into the model’s input embedding space.
  3. The model reasons over it with text. The projected data is joined with your normal text prompt, so you can ask questions like “predict the band gap of this crystal” or “where is the splice site in this sequence”.

The model card lists these input families:

  • Chemistry and materials: molecular graphs, crystal structures (CIF files), mass spectra, X-ray diffraction.
  • Biology: protein sequences and 3D structures, DNA and RNA sequences.
  • Medical signals: ECG and EEG.
  • Astronomy: light curves.

The tasks named on the card are materials property prediction, sequence classification, site localisation and structure inference. The first public projector folders cover crystals (JARVIS), DNA splice sites, ECG, EEG and protein binding.

Warning: The ECG and EEG adapters are research tools for specific benchmark tasks. They are not medical devices, so never use the output to diagnose a person.

All Benchmarks (Vendor-Reported)

Shanghai AI Lab published 33 benchmarks across seven areas. Every number below is vendor-reported; nobody has reproduced them independently yet, and the lab has not published its test harness or logs.

The comparison models are exactly the ones the lab chose. “–” means the lab gave no score.

Agentic coding and productivity

Benchmark Sage S2.1 Flash GLM-5.3-Flash DeepSeek-V4.1-Flash Qwen3.8-Max GPT-5.6 Sol Claude Opus 5
Terminal-Bench 2.1 87.9 84.3 87.6 89.6 86.2 90.2
DeepSWE 67.9 63.7 73.9 67.0 70.6 72.6
NL2Repo-Bench 57.2 53.0 62.7 63.2 60.7 79.9
SWE-bench Pro 60.6 60.3 61.8 65.3 59.8 76.5
SpreadsheetBench 2 37.1 32.4 35.8 34.9 32.7 33.6
ALE-CLI 32.4 28.1 35.4 24.8 28.6 29.5
APEX-Agents 61.7 50.8 69.2 66.3 55.4 61.3
GDPval 78.0 67.7 77.6 79.6 69.6 80.0
Workspace-Bench-Lite 67.1 65.1 66.8 67.4 57.4 68.1
Workspace-Bench 61.7 59.7 63.0 63.9 50.9 62.0

Tool use, ML engineering and research

Benchmark Sage S2.1 Flash GLM-5.3-Flash DeepSeek-V4.1-Flash Qwen3.8-Max GPT-5.6 Sol Claude Opus 5
OfficeQA Pro 72.2 64.7 69.7 66.2 66.4 67.2
Harvey LAB 25.0 20.0 11.7 24.2 5.0 22.7
AutomationBench 56.7 48.3 57.8 49.7 45.7 49.4
BFCL V4 75.5 71.7 71.4 66.8 68.8 80.6
JobBench 58.8 56.5 58.5 52.7 45.4 68.0
SkillsBench 62.3 58.3 52.8 66.7 62.5 63.7
τ³-bench Banking 45.4 39.2 47.4 55.2 46.9 48.7
MLE-bench Lite 90.3 84.3 89.2 82.2 80.7 88.0
PostTrainBench 39.2 36.8 – – 31.0 35.7
PaperBench 88.0 83.0 88.6 95.1 94.3 86.9
MLS-Bench 42.0 39.9 44.3 41.0 46.2 49.8
DeepResearch Bench II 54.2 51.9 55.1 49.2 50.7 54.1

Search and vision

Benchmark Sage S2.1 Flash GLM-5.3-Flash DeepSeek-V4.1-Flash Qwen3.8-Max GPT-5.6 Sol Claude Opus 5
DeepSearchQA 94.4 92.6 95.3 – 93.2 –
BrowseComp-VL 60.5 61.0 60.6 – – –
WideSearch 81.5 77.2 86.3 81.9 83.3 84.0
BrowseComp 91.2 88.3 93.8 – 92.2 90.8
OSWorld v2.1 69.3 66.4 70.1 66.4 70.4 74.0
OSWorld-Verified 85.2 83.5 78.5 84.3 82.2 86.4
WebArena-Verified 67.6 65.0 60.7 64.9 66.6 69.1
Vision2Web 78.0 75.8 – 75.2 73.5 80.2
WeaveBench 68.4 60.9 – 71.9 79.6 86.4
OSWorld-G 84.8 83.7 61.4 88.6 77.7 76.8
BabyVision 75.8 51.3 58.0 83.8 80.7 63.1

How to read these numbers

  • Clear wins: office and spreadsheet work (OfficeQA Pro, SpreadsheetBench 2), legal tasks (Harvey LAB) and ML engineering (MLE-bench Lite, PostTrainBench).
  • Big jump over its base: BabyVision 75.8 vs 51.3 for GLM-5.3-Flash, so the extra vision training clearly did something.
  • Weak spot: hard software engineering. Claude Opus 5 leads SWE-bench Pro by about 16 points and NL2Repo by about 23.
  • Missing: the table has no science-specific benchmarks with numbers for the projector tasks, which is odd for a “scientific” model. Those results may be in the full paper.

Sage S2.1 Flash vs GLM-5.3-Flash, DeepSeek V4.1 Flash and Qwen3.8

Here is how Sage sits next to the open models you are most likely to compare it with. Scores come from the Sage table above; other facts come from each model’s own release.

Model Size Licence Special strength Runs at home?
Sage S2.1 Flash 320B, ~18B active MIT Scientific data inputs, office and ML-engineering agents No (330 GB even in FP8)
GLM-5.3-Flash (Z.ai) 320B, ~18B active Open weights The base architecture; general chat and coding Only with huge RAM (see our GLM Flash local guide)
DeepSeek V4.1 Flash MoE Open weights Best search and many agent scores in this table Yes with big RAM and GGUF (see our guide)
Qwen3.8-Flash-Next 125B, 6B active Qwen licence Fast, multimodal, runs on 2x 16 GB GPUs + 128 GB RAM Yes (see our guide)
LongCat-2.5 (Meituan) 1.6T, ~48B active MIT Vision and video, cheap API No (see our guide)

Who should pick which:

  • Labs working with molecules, crystals, sequences or signals: Sage is the only open model here with native projectors for that data.
  • Coding agents: DeepSeek V4.1 Flash or Qwen3.8-Max beat Sage on SWE-bench Pro and DeepSWE, and Claude Opus 5 is far ahead if you can use a closed API.
  • Home users with one or two GPUs: Qwen3.8-Flash-Next is the realistic choice.

Hardware: What You Need to Run It

Version Disk size Vendor example Notes
BF16 (Sage-S2.1-Flash) ~645 GB 8x H200 (141 GB each) Tensor parallel 8 in vLLM and SGLang
FP8 (Sage-S2.1-Flash-FP8) ~330 GB 4x H200 Tensor parallel 4 in vLLM and SGLang; the LMDeploy example still uses 8
GGUF / 4-bit Not official – The Hub lists community quantizations, but none is official or tested by the lab

Can you run it on a gaming PC? Realistically, no. Even a rough 4-bit file would be around 170-190 GB (my estimate from 320B parameters), which is more than a 2x RTX 5080 rig with 128 GB RAM can hold.

llama.cpp support for the glm5_next architecture also has to be in place before any GGUF can work. Check the model page’s quantization list before downloading a community file.

Tip: For a short test, rent a 4x H200 node for the FP8 version instead of buying hardware. Compare hourly rates in our cheapest cloud GPU guide, and make sure the node has at least 400 GB of fast disk for the download.

How to Run It: vLLM, SGLang and LMDeploy

The official deployment guide gives commands for three engines. All three expose an OpenAI-compatible API. If you are new to these engines, our vLLM vs llama.cpp vs Ollama comparison explains the differences.

Warning: Always pass --trust-remote-code. Without it, the engine silently loads a stock tokenizer that mis-reads tool-call tags, so tool calls and reasoning break with no error message.

Option 1: vLLM 0.33.0 (with the patch)

  1. Install vLLM 0.33.0 in a fresh Python environment.
  2. Download the patch folder from the model repo:
    hf download internlm/Sage-S2.1-Flash --include "vllm_patch/*" --local-dir /path/to/Sage-S2.1-Flash

    Apply it as its README says. It changes vllm/renderers/hf.py so images returned by parallel tool calls attach to the right result.

  3. Start the server (BF16, 8 GPUs):
    vllm serve internlm/Sage-S2.1-Flash --trust-remote-code --tensor-parallel-size 8 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm47
  4. Or the FP8 version on 4 GPUs:
    vllm serve internlm/Sage-S2.1-Flash-FP8 --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm47

Use the default vLLM frontend. The faster Rust frontend from the GLM recipes does not accept image or video input with this model yet.

Option 2: SGLang 0.5.20 or later

  1. Download the chat template (SGLang otherwise uses its own built-in one):
    hf download internlm/Sage-S2.1-Flash chat_template.jinja --local-dir /path/to/Sage-S2.1-Flash
  2. Launch (BF16):
    python3 -m sglang.launch_server --model-path internlm/Sage-S2.1-Flash --trust-remote-code --chat-template /path/to/Sage-S2.1-Flash/chat_template.jinja --tp-size 8 --reasoning-parser glm45 --tool-call-parser glm47
  3. For FP8, swap in internlm/Sage-S2.1-Flash-FP8, its own template path, and --tp-size 4.

Option 3: LMDeploy (main branch)

  1. Pin transformers to a supported range:
    pip install "transformers>=5.17,<5.19"

    Version 5.19 renames an attention class that LMDeploy does not accept yet. FlashMLA is also required.

  2. Serve the model:
    lmdeploy serve api_server internlm/Sage-S2.1-Flash --trust-remote-code --backend pytorch --tp 8 --reasoning-parser default --tool-call-parser glm47

Send your first request

Once any server is up (vLLM listens on port 8000 by default), call it like the OpenAI API:

curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{"model": "internlm/Sage-S2.1-Flash", "messages": [{"role": "user", "content": "Explain what an XRD pattern tells a materials scientist."}]}'

The generation config in the repo sets temperature 1.0 and top_p 0.95, so leave those as defaults unless you have a reason to change them.

How to Use the Scientific Projectors

Scientific inputs do not go through the normal chat endpoint. You compute embeddings yourself and send them as prompt_embeds. Here are the official steps.

  1. Install the basics:
    python -m pip install torch safetensors huggingface_hub
  2. Download the projectors:
    hf download internlm/Sage-S2.1-Flash --include "projectors/sage-s2.1-flash/**" --local-dir ./sage-s2.1-flash
  3. Download the loader scripts from the Sage-SciPerception repo:
    hf download internlm/Sage-SciPerception load_projector.py modeling_projector.py --local-dir ./sage-s2.1-flash
  4. Test-load a projector on CPU (this one is for crystals):
    cd sage-s2.1-flash
    python -c "from load_projector import load_projector; p, c = load_projector('projectors/sage-s2.1-flash/cif/jarvis', device='cpu'); print(c)"
  5. Start vLLM with prompt embeddings on. Add --enable-prompt-embeds to the vLLM serve command.
  6. Build and send the input. Run the matching encoder on your data, pass it through the projector, join the result with the text token embeddings, and send it to /v1/completions as prompt_embeds.

Note: Each task’s encoder, readout and preprocessing are set in that projector’s config.json. The Sage-SciPerception model page has a full CIF (crystal) client example, which is the easiest place to start.

Limits, Licence and Open Questions

Licence

  • MIT for code and weights, per both model cards. You can use, modify and sell products built on it; keep the licence notice.
  • The arXiv paper itself uses the standard arXiv licence, which is separate from the weights.

Known limits

  • All scores are vendor-reported. No harness, logs or per-task results are public, and the model is not on Artificial Analysis yet.
  • Coding gap: 60.6 on SWE-bench Pro vs 76.5 for Claude Opus 5.
  • No official context window. The config allows about 1M positions, but nobody has shown long-context quality yet.
  • Audio unclear: the chat template has audio slots but there is no audio config.
  • No hosted API with pricing. A “try the API” link on the website points to discovery.intern-ai.org.cn, but access terms are not clear.
  • Documentation gaps: the model card mentions “15 benchmarks” in one place and shows 33 in the table, and the paper abstract uses the old S1.5 name.

Privacy and safety

Because you host the weights yourself, your data never leaves your servers. That makes it attractive for labs with sensitive research data. Still, treat its scientific predictions as hypotheses to check in the lab, not as results.

Troubleshooting

  • Tool calls come back as plain text or reasoning gets cut off: you probably forgot --trust-remote-code. Add it and restart.
  • SGLang ignores tools or formats replies oddly: pass the repo’s chat_template.jinja with --chat-template.
  • LMDeploy fails on indexed_attention: you have transformers 5.19 or newer. Reinstall with "transformers>=5.17,<5.19".
  • Images from parallel tool calls land on the wrong result in vLLM: apply the repo’s vllm_patch. The upstream fix is vLLM pull request 67108.
  • A non-streaming reply swallows a tool call or ends early: this is known vLLM issue 64326, triggered by a literal closing think tag. Use streaming responses until it is fixed.
  • Crash or garbage output with FP8 KV cache on H100 or H200: FP8 KV cache is not supported for this model on Hopper. Keep the default BF16 KV cache.
  • Out of memory on 4x H100 (80 GB): 320 GB of 330 GB FP8 weights leaves no room for the KV cache. Use 4x H200, 8x H100, or larger.

My Take

Sage S2.1 Flash is not a model most of our readers will run at home, and that is fine. Its real story is that a big lab released a frontier-class, MIT-licensed agent model that natively reads scientific data.

With most open chat models, you would have to bolt a separate chemistry or biology model on yourself.

The benchmark table is honest in one useful way: it shows where Sage loses. It is not the best coder, and Claude Opus 5 still leads most agent tests. Where it shines is the “boring but valuable” work: spreadsheets, office documents, legal tasks and ML experiments.

I would wait for two things before betting on it: independent benchmark runs, and published numbers for the science projector tasks. If you are a research team with H200 access, the FP8 version on 4 GPUs is cheap enough to test this week.

I will update this guide when an API price, independent scores or an official quantized build appears.

FAQ

Who made Sage S2.1 Flash?

Shanghai Artificial Intelligence Laboratory, the team behind InternLM. The weights are under the internlm organisation on Hugging Face.

Is Sage S2.1 Flash free for commercial use?

Yes. Both the code and the weights are under the MIT licence, which allows commercial use as long as you keep the licence notice.

How much VRAM does Sage S2.1 Flash need?

The FP8 weights are about 330 GB and the vendor runs them on 4x H200. BF16 is about 645 GB on 8x H200. It will not fit on consumer GPUs.

Is there a GGUF or Ollama version?

Not an official one. The Hub lists community quantizations, but llama.cpp support for its glm5_next architecture and the quality of those files are unverified.

Is it better than Claude Opus 5?

On the lab’s own numbers it beats Opus 5 on MLE-bench Lite, OfficeQA Pro, SpreadsheetBench 2 and Harvey LAB, but loses clearly on SWE-bench Pro, NL2Repo and most vision-agent tests.

What scientific data can it read?

Molecular graphs, crystal structures, protein sequences and structures, DNA and RNA, mass spectra, X-ray diffraction, light curves, ECG and EEG, through small projector add-ons.

Is there an API I can call?

Not with public pricing yet. You can self-host it with vLLM, SGLang or LMDeploy, which all give an OpenAI-compatible endpoint.

What is the difference between Sage S2.1 Flash and GLM-5.3-Flash?

Sage uses the GLM-5.3-Flash architecture and size, but adds scientific projectors and agent training. On the lab’s table it scores higher than GLM-5.3-Flash on almost every benchmark, most visibly BabyVision (75.8 vs 51.3).

Sources

  • Hugging Face model card: internlm/Sage-S2.1-Flash (licence, benchmark table, SciPerception steps)
  • Hugging Face: Sage-S2.1-Flash deployment_guide.md (vLLM, SGLang, LMDeploy commands and versions)
  • Hugging Face model card and deployment guide: internlm/Sage-S2.1-Flash-FP8
  • arXiv 2611.21072: “Sage S2.1 Flash Technical Report”, submitted 16 Nov 2026
  • Hugging Face: internlm/Sage-SciPerception (projector loader scripts)
  • Official website: internlm.github.io/Sage-S2.1-Flash
  • ModelScope: Shanghai_AI_Laboratory/Sage-S2.1-Flash
  • OrcaRouter: “Sage S2.1 Flash quiet release” (repo file and config analysis)
  • Let’s Data Science: “Sage S2.1 Flash Report Details 320B Scientific Reasoning Model”, 17 Nov 2026
  • ai-tldr.dev release notes and llmreference.com model page
Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Woman in a science lab holding a tablet with a SAGE card, hook 320B open science AI