Shanghai Artificial Intelligence Laboratory · Released, not on our provider
Sage S2.1 Flash (2026): what it is, its benchmarks, the hardware it needs and how to run it
Sage S2.1 Flash is an open-weight scientific multimodal model from Shanghai Artificial Intelligence Laboratory, a 320B-parameter mixture-of-experts with 18B active parameters and a 1M-token configured context, published under the MIT license on Hugging Face in November 2026 for multi-step research, coding, office and tool-use work.
Use it now
Chat
This chat starts with DeepSeek V4.1 Flash, not Sage S2.1 Flash.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- Shanghai Artificial Intelligence Laboratory
- Type
- Text
- Released
- No date given
- OpenRouter id
- None
Checked
Sage S2.1 Flash: facts, with sources
- Hugging Face’s API lists the repository internlm/Sage-S2.1-Flash as created on November 15, 2026 and last modified on November 16. The model card prints no release date, so we give the repository creation date and do not call it a release date. The lab’s card describes the model as a scientific multimodal model from Shanghai Artificial Intelligence Laboratory.Source: huggingface.co · checked 2026-11-19
- The card says the model is designed for research tasks that need several rounds of analysis and verification: it combines scientific data understanding with information retrieval, programming and tool use, can run specialized software for analysis, and checks results against intermediate outputs and visual feedback before adjusting later steps.Source: huggingface.co · checked 2026-11-19
- Through its Sage SciPerception architecture the card says the model accepts more than ten types of scientific input, including molecular graphs, crystal structures, and protein sequences and 3D structures. Each input type is encoded and projected into a shared space and read next to the text prompt, for prediction, classification, site localization, quantitative analysis and structure inference.Source: huggingface.co · checked 2026-11-19
- The repository’s deployment guide calls the release a 320B-A18B multimodal mixture-of-experts model (320 billion parameters in total, 18 billion active), stored in bfloat16 at about 645 GB, built on the GLM-5.3-Flash architecture. Hugging Face’s own count for the repository is 321,323,031,390 parameters.Source: huggingface.co · checked 2026-11-19
- The repository’s config.json lists 45 layers, 288 routed experts plus 1 shared expert with 8 experts used per token, and a maximum position length of 1,048,576 tokens (about 1M). The card text itself does not state a context length.Source: huggingface.co · checked 2026-11-19
- The card says the code and model weights in the repository are released under the MIT License. The weights are not gated on Hugging Face.Source: huggingface.co · checked 2026-11-19
- An arXiv paper titled "Sage S2.1 Flash Technical Report" (arXiv:2611.21072, submitted November 16, 2026, revised November 17) describes a 320-billion-parameter general-purpose mixture-of-experts model and says it was evaluated on 33 benchmarks across seven domains. The model card we read does not link to it and we did not confirm the authors’ affiliation, so we treat it as a second description, not as the maker’s page.Source: arxiv.org · checked 2026-11-19 · second-hand
- Sage S2.1 Flash is not on OpenRouter’s model list (checked November 19, 2026), so this site cannot run it. This site is a text chat with other models, such as DeepSeek V4.1 Flash and Qwen3.8-Max.Source: openrouter.ai · checked 2026-11-19 · second-hand
Our table
Sage S2.1 Flash benchmarks: the lab’s own table against five other models
- The card’s text says Sage S2.1 Flash placed first on MLE-bench Lite with 90.3, ahead of DeepSeek-V4.1-Flash-1017 at 89.2, and first on OfficeQA Pro (72.2), Harvey LAB (25.0) and SpreadsheetBench 2 (37.1). In the card’s table its number is also the highest on PostTrainBench (39.2). These are the lab’s own figures, not tested by us.Source: huggingface.co · checked 2026-11-19
- The same table shows it behind other models on several coding and agent rows: SWE-bench Pro 60.6 against 76.5 for Claude Opus 5, NL2Repo-Bench 57.2 against 79.9, Terminal-Bench 2.1 87.9 against 90.2, and DeepSWE 67.9 against 73.9 for DeepSeek-V4.1-Flash-1017. The card’s text counts first place on four benchmarks and second place on seven of the 15 in its chart; the table below the chart lists 33 rows.Source: huggingface.co · checked 2026-11-19
Vendor benchmarks
Sage S2.1 Flash against GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Qwen3.8-Max, GPT-5.6 Sol and Claude Opus 5
People who search for Sage S2.1 Flash want to know whether an open 320B model stands next to the big closed ones. The lab’s table gives a mixed answer. Sage is the top number on office documents, spreadsheets and machine-learning engineering tasks, but Claude Opus 5 is clearly ahead on SWE-bench Pro and BFCL V4, and the card shows other models ahead on several search and vision rows (BrowseComp, BabyVision). The card does not say how each competitor score was produced in the text we read, so use it as a guide to where the lab says it is strong, not as a ranking. If you want a model you can chat with today, DeepSeek V4.1 Flash and Qwen3.8 run on this site and Kimi K3 is another large model with its own page.
| Benchmark | Sage S2.1 Flash | GLM-5.3-Flash | DeepSeek-V4.1-Flash-1017 | 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 |
| SWE-bench Pro | 60.6 | 60.3 | 61.8 | 65.3 | 59.8 | 76.5 |
| DeepSWE | 67.9 | 63.7 | 73.9 | 67.0 | 70.6 | 72.6 |
| SpreadsheetBench 2 | 37.1 | 32.4 | 35.8 | 34.9 | 32.7 | 33.6 |
| GDPval | 78.0 | 67.7 | 77.6 | 79.6 | 69.6 | 80.0 |
| 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 |
| BFCL V4 | 75.5 | 71.7 | 71.4 | 66.8 | 68.8 | 80.6 |
| MLE-bench Lite | 90.3 | 84.3 | 89.2 | 82.2 | 80.7 | 88.0 |
| PostTrainBench | 39.2 | 36.8 | – | – | 31.0 | 35.7 |
| BrowseComp | 91.2 | 88.3 | 93.8 | – | 92.2 | 90.8 |
| OSWorld-Verified | 85.2 | 83.5 | 78.5 | 84.3 | 82.2 | 86.4 |
| BabyVision | 75.8 | 51.3 | 58.0 | 83.8 | 80.7 | 63.1 |
Our table
How to run Sage S2.1 Flash: the weights, the engines and what each needs
- The deployment guide’s vLLM and SGLang examples use 8× H200 GPUs (tensor parallel 8), and the LMDeploy example also uses tensor parallel 8. Every engine needs --trust-remote-code, because the repository ships its own tokenizer code; without it the engines silently load a stock tokenizer and tool calls and reasoning can break, according to the guide.Source: huggingface.co · checked 2026-11-19
- Version notes in the guide: vLLM 0.33.0 plus a patch file shipped in the repository (it fixes images being attached to the wrong tool result), SGLang 0.5.20 or later with the repository’s chat template, and LMDeploy built from main with transformers at least 5.17 and below 5.19 plus FlashMLA. On Hopper GPUs (H100, H200) keep the default BF16 KV cache, since an FP8 KV cache is not supported for this model there.Source: huggingface.co · checked 2026-11-19
- A separate repository, internlm/Sage-S2.1-Flash-FP8, holds an FP8 copy. Hugging Face’s API shows it was created on November 13, 2026, that most of its parameters are stored as 8-bit floats, and that its files add up to about 329 GB, roughly half of the BF16 repository. The introduction on the FP8 card repeats the BF16 card’s text, and the card points to its own deployment guide.Source: huggingface.co · checked 2026-11-19
- For scientific inputs the card says to download the projector from the model repository and the loading scripts from the Sage-SciPerception repository, run the scientific encoders yourself, and send the combined embeddings to a vLLM server started with --enable-prompt-embeds through /v1/completions. A first check can load the projector on the CPU.Source: huggingface.co · checked 2026-11-19
Run it yourself
Ways to run Sage S2.1 Flash and what each one needs
This is not a model you load on a laptop. The lab’s guide sizes the BF16 weights at about 645 GB and shows 8× H200 GPUs for its vLLM and SGLang examples, so in practice it is for a lab or a company with a GPU server, or for a hosting provider. Because the weights are MIT-licensed, a provider can host it and sell access, which is the usual way a model like this reaches an API list; none has appeared on OpenRouter as of our check. Until then, a hosted model such as DeepSeek V4.1 Flash is the practical way to try this kind of agent work without owning the hardware.
| Route | What it is | What it needs |
|---|---|---|
| BF16 weights | The main repository, MIT license, 321.3 billion parameters by Hugging Face’s count | About 645 GB of storage; the guide’s serving examples use 8× H200 |
| FP8 weights | A separate repository with 8-bit float weights and its own deployment guide | Files add up to about 329 GB; check the FP8 guide for its engine settings |
| vLLM | Serve with --trust-remote-code, tool and reasoning parsers glm47, tensor parallel 8 | vLLM 0.33.0 plus the repository’s patch; default Python frontend (the Rust frontend takes no image or video input yet) |
| SGLang | Serve with --trust-remote-code and the repository’s own chat template, tensor parallel 8 | SGLang 0.5.20 or later; download chat_template.jinja first |
| LMDeploy | PyTorch backend, tensor parallel 8, tool parser glm47 | LMDeploy built from main, transformers from 5.17 up to but not including 5.19, FlashMLA installed |
| Scientific inputs (projectors) | Projector from the model repository, loader scripts from Sage-SciPerception, prompt embeddings sent to /v1/completions | A vLLM server with --enable-prompt-embeds, plus the scientific encoders and preprocessing set up separately |
Source: huggingface.coSource: huggingface.coSource: huggingface.co
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to Sage S2.1 Flash you can use now
DeepSeek V4.1 Flash
DeepSeek$0.30 / $1.2 (per million tokens)
- DeepSeek released V4.1-Flash on September 10, 2026.Source: api-docs.deepseek.com · checked 2026-11-11
- Only the Flash version of V4.1 exists: DeepSeek has published no V4.1 Pro, and the Pro model on its price page is still deepseek-v4-pro (V4-Pro-0813).Source: api-docs.deepseek.com · checked 2026-11-11
- DeepSeek’s pricing page says the API model deepseek-flash is now served by the V4.1-Flash-1017 snapshot.Source: api-docs.deepseek.com · checked 2026-11-11
Qwen3.8-Max
Alibaba$2 / $6 (per million tokens)
- Alibaba’s price for qwen3.8-max is $2 / $6 per million input / output tokens (Singapore region).Source: docs.qwencloud.com · checked 2026-11-13
- Alibaba describes qwen3.8-max as a 2.4-trillion-parameter mixture-of-experts model.Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-max-0902 at $2 / $6 per million tokens.Source: openrouter.ai · checked 2026-11-13
FAQ
Sage S2.1 Flash FAQ
What is Sage S2.1 Flash?
Sage S2.1 Flash is a scientific multimodal model from Shanghai Artificial Intelligence Laboratory, published as open weights in the internlm organization on Hugging Face. The lab says it is built for research tasks that need several rounds of analysis and verification, combining scientific data understanding with retrieval, programming and tool use. The deployment guide calls it a 320B-A18B mixture-of-experts model on the GLM-5.3-Flash architecture.
Who made Sage S2.1 Flash and when did it come out?
The card names Shanghai Artificial Intelligence Laboratory, and the weights sit in the internlm organization on Hugging Face (the Chinese name on the card is 书生 S2.1 Flash). The card prints no release date. Hugging Face shows the main repository was created on November 15, 2026 and the FP8 repository on November 13, and an arXiv technical report was submitted on November 16.
How big is Sage S2.1 Flash and what hardware does it need?
It has about 321 billion parameters by Hugging Face’s count, of which 18 billion are active per token, and the BF16 weights are about 645 GB. The lab’s vLLM and SGLang examples use 8× H200 GPUs. An FP8 copy of about 329 GB exists, so it is a server-class model, not one for a laptop or a single consumer GPU.
What license does Sage S2.1 Flash use?
The card says the code and the model weights are released under the MIT License, which is a permissive license that allows use, copying, modification and redistribution with the license text kept. Read the LICENSE file for the exact terms. Note that the engines also need the repository’s custom code, loaded with trust_remote_code, so review it before running it on your own servers.
How good is Sage S2.1 Flash?
By the lab’s own table, which we did not run, it scores 90.3 on MLE-bench Lite, 72.2 on OfficeQA Pro and 37.1 on SpreadsheetBench 2, the top numbers in that table, but 60.6 on SWE-bench Pro against 76.5 for Claude Opus 5. It is strongest on documents, spreadsheets and machine-learning tasks and weaker on software-engineering rows. The scores are from the maker, not independent tests.
What can it do with scientific data?
The card says the Sage SciPerception architecture lets it read more than ten kinds of scientific input, such as molecular graphs, crystal structures and protein sequences, next to a text prompt, for prediction, classification and structure inference. This needs a separate projector, the encoders and a vLLM server with prompt embeddings, so it is a research setup. The lab gives no accuracy numbers for these tasks on the model card.
Can I chat with Sage S2.1 Flash here, and what can I use instead?
No. It is not on OpenRouter’s model list, so this site cannot run it. For hosted agent-style work you can chat with DeepSeek V4.1 Flash: on this site a typical reply costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Qwen3.8 is another option.
Is this the official Sage S2.1 Flash website?
No. This page is unofficial and not affiliated with or endorsed by Shanghai Artificial Intelligence Laboratory. Facts link to the pages we read.