A generated title card reading 'Sage-S2.1-Flash' above the subtitle '645 GB of weights, a 21-byte model card, no announcement', with three small chips below reading '15 Nov 2026 upload', 'MIT licence' and '320B-A18B MoE'. White background, blue and cyan gradient accents, flat line icons. The OrcaRouter logo is composited in the bottom-right corner.
Guides & Insights

Sage-S2.1-Flash Shipped Quietly: 645 GB of Weights, a 21-Byte Model Card, and a Benchmark Table Nobody Linked

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On 15 November 2026 at 02:34 UTC, an account belonging to the Shanghai AI Laboratory uploaded a repository called internlm/Sage-S2.1-Flash to Hugging Face. It contained 120 safetensors shards totalling about 645 gigabytes, a configuration file, a chat template, and a README of 21 bytes. There was no blog post. The lab's own homepage carried no mention of the name when we fetched it on 16 November. There was no press release that we could find, and no entry for Sage-S2.1-Flash on Artificial Analysis.

What did exist, by the end of that weekend, was three things a reader can check: the weights themselves, a deployment guide written for people who intend to serve the thing on eight H200s, and a benchmark table hosted on a GitHub Pages microsite that the model repository never links to. This piece is an inventory of those three, with a clear line between what the files prove and what InternLM has not yet said.

Screenshot of the Hugging Face 'Files and versions' tab for internlm/Sage-S2.1-Flash. The header shows the safetensors and glm5_next tags, the MIT licence, and a repository size of 643 GB. The file list shows .gitattributes, README.md, chat_template.jinja, config.json, deployment_guide.md, generation_config.json, and shards numbered model-00001-of-00120 through model-00013-of-00120 and beyond.

The repository, file by file

The evidence is unusually clean because there is so little of it. A directory listing on 16 November returns 155 files. One hundred and twenty of them are weight shards of roughly 5.35 GB each. The rest are the machinery of a multimodal model: tokenizer.json at 20 megabytes, a 10.5 KB chat_template.jinja, a 5.9 KB config.json, a generation config with temperature 1.0 and top_p 0.95, a deployment_guide.md, and two directories — one holding a vLLM patch, one holding projectors for a companion perception model.

The README is worth dwelling on. Twenty-one bytes, and every one of them is YAML frontmatter:

• A horizontal rule, then one line reading license: mit, then another horizontal rule.

That is the whole model card. There is no description, no benchmark table, no usage snippet, no citation block, no link to a paper. Hugging Face renders it as a page with a title, an MIT badge and an empty body. The repository has one like and, at the time of writing, zero downloads — which for a model this size is less a signal about quality than about how few people can afford the disk.

Two smaller details from the file listing are worth recording because they date the release better than any announcement would. The FP8-quantised sibling, internlm/Sage-S2.1-Flash-FP8, was created on 13 November at 15:04 UTC — roughly 35 hours before the bfloat16 repository. And the bf16 repo's last commit, at 03:01 UTC on 16 November, is a bare "Upload folder using huggingface_hub", the same message as the initial upload. Whatever workflow produced this, it was not a curated launch.

The one document the lab did write

deployment_guide.md is the only prose InternLM shipped inside the repo, and it is the most informative file there. Its opening sentence is the whole specification:

"The Sage-S2.1-Flash release is a 320B-A18B multimodal MoE model stored in bfloat16 weight format (about 645 GB)."

Three numbers, and the Hub's own metadata confirms two of them. The parameter count on the model page reads 321 billion, a difference of one billion that is almost certainly a rounding convention rather than a disagreement. The storage figure checks out against internlm's repository metadata at 642.8 GB. "A18B" — 18 billion active parameters per token — is InternLM's own term and is not independently derivable from the config, so treat it as a vendor figure.

The config.json fills in the rest, and it is a genuinely interesting document. The model has 45 layers, and only 11 of them use conventional sparse attention. The other 34 use a linear-attention layer type that the config calls kda, keyed by a per-layer list in linear_attn_config. Routing is 288 experts with one shared expert and eight selected per token, at a routed scaling factor of 2.5. The positional ceiling is 1,048,576 tokens. There is a vision tower — depth 24, 448-pixel images, patch size 14 — and the token vocabulary carries explicit image, video, image-start, image-end, video-start and video-end tokens.

Screenshot of deployment_guide.md rendered in the Hugging Face file viewer for internlm/Sage-S2.1-Flash. The opening sentence reads 'The Sage-S2.1-Flash release is a 320B-A18B multimodal MoE model stored in bfloat16 weight format (about 645 GB).' Below it, the guide states the vLLM and SGLang examples use 8x H200 GPUs and that the model uses the GLM-5.3-Flash architecture, followed by the 'Required for Sage-S2.1-Flash' section on --trust-remote-code and the vLLM patch.

The architecture line that matters most

The deployment guide contains one sentence that reframes the whole release. It says Sage-S2.1-Flash "uses the GLM-5.3-Flash architecture", and instructs the reader to install and tune vLLM and SGLang by following Z.ai's official GLM-5.3-Flash recipes, substituting one repository name for the other.

That is confirmed in the config: the architecture class is Glm5NextForConditionalGeneration and the model type is glm5_next. So a Shanghai AI Laboratory model is published on top of a Z.ai architecture, and the lab says so plainly rather than hiding it. GLM-5.3-Flash is itself a 320-billion-parameter, 18-billion-active multimodal MoE with a 1M-token context and a 128K output ceiling — which is to say the parameter shape is not a coincidence either.

For anyone deciding whether to deploy this, that sentence is the single most valuable thing in the release. It means the serving story is not new. It means the recipes, the kernel work and the known failure modes are the ones already documented for a model that has been public since August. It also means the patches InternLM added are worth reading closely, because they describe exactly where their model diverges from the base.

Three ways this release breaks if you serve it naively

The deployment guide is, read carefully, a list of traps. Each one is a silent failure rather than an error, which is the dangerous kind.

• Missing --trust-remote-code breaks tool calls without warning. Agentic traces contain strings like <|user|> and <tool_call> as ordinary text. The repository ships its own tokenization code that treats them as text; without the flag, the engine loads the stock tokenizer, encodes them as control tokens, and the model can stop early or corrupt a tool call. Nothing logs.

• vLLM needs a one-file patch, or images get attached to the wrong tool result. When parallel tool calls return in a different order than they were issued, the chat template renders results in call order while vLLM collects images in message order. Each image silently lands on the wrong result and the model answers about the wrong picture. The fix replaces vllm/renderers/hf.py; the upstream fix is in flight as vllm-project/vllm#67108.

• SGLang will not use the repository's chat template unless you hand it the file. SGLang installs its own built-in template for this architecture by default, which bypasses the repository's tokenizer entirely; you have to download chat_template.jinja and pass it with --chat-template.

There are two more constraints in the same document. FP8 KV cache is not supported on Hopper GPUs, so on H100s and H200s you keep the BF16 cache. And LMDeploy needs transformers between 5.17 and 5.19, because 5.19 renames the layer type deepseek_sparse_attention to indexed_attention and LMDeploy rejects the new name. The guide also flags an open vLLM issue — vllm-project/vllm#64326 — where a literal </think> in non-streaming responses can truncate reasoning or swallow a tool call; streaming is unaffected.

None of this says the model is bad. It says the release is at the stage where the people who wrote it are still the only people who have run it, and they are telling you so.

The benchmark table that is not in the repository

InternLM did publish evaluation numbers — 35 of them, across Terminal-Bench 2.1, DeepSWE, SWE-bench Pro, APEX-Agents, GDPval, OfficeQA Pro V2, Harvey LAB, OSWorld-Verified, BrowseComp, BabyVision and two dozen more. They are not in the model card, because the model card is empty. They live in a separate GitHub repository, InternLM/Sage-S2.1-Flash, created on 16 November at 03:34 UTC, described as "Sage 2.1 Flash bilingual website and application demos", which builds a static site at internlm.github.io/Sage-S2.1-Flash. The benchmark data is committed there as a tab-separated file and a JSON file, alongside demo videos of the model driving 3D Slicer, Blender and Fiji.

That site's English copy opens with "Today, we introduce Sage 2.1 Flash, a frontier scientific multimodal agent model". So the announcement does exist — it is just on a GitHub Pages host that nothing in the Hugging Face repository points to, and it was created after the weights were already downloadable. If you want the vendor's own framing, that paragraph is it: more than ten kinds of scientific input in one model, including molecular graphs, crystal structures, protein sequences and 3D structures, DNA and RNA, spectra, mass spectra, X-ray diffraction, electrocardiograms, electroencephalograms and light curves.

The evaluation table compares Sage 2.1 Flash against eight models: GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Qwen3.8-Max, GPT-5.6 Sol, Claude Opus 5, GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5. Every number in it is vendor-reported. There is no published harness, no run log, no per-task breakdown, and no third party has reproduced any of it. Read the table as a statement of intent by the lab, not as measurement.

Where the vendor table is unusually strong, and unusually weak

Two patterns stand out, and both are more informative than the headline scores.

The first is that the scientific and vision rows are where the margins are absurd. BabyVision reads 75.8 for Sage against 51.3 for GLM-5.3-Flash, 58.0 for DeepSeek-V4.1-Flash-1017 and 63.1 for Claude Opus 5. OSWorld-G reads 84.8 against 61.4 for DeepSeek. OfficeQA Pro V2 reads 72.2, the highest in its own table, ahead of Claude Opus 5 at 67.2. When a model wins a row by twenty-four points, either the capability is real or the benchmark is being read differently for this model than for the others — and with no harness published, you cannot tell which.

The second is that the coding rows are the least flattering document in the release. Terminal-Bench 2.1 at 87.9 is strong but not leading; Qwen3.8-Max is at 89.6 and Claude Opus 5 at 90.2. SWE-bench Pro at 60.6 trails DeepSeek-V4.1-Flash-1017 at 64.7 and Qwen3.8-Max at 65.3. DeepSWE at 67.0 trails DeepSeek at 73.9 by nearly seven points. For a model whose marketing is agentic, that is a table its own authors chose to publish, which is worth something.

What the table does not contain is as notable: Claude Haiku 5.5 and Gemini 3.8 Flash are both absent, so the two comparisons most people would want cannot be made from vendor data at all, in either direction.

A generated infographic titled 'Sage-S2.1-Flash - what the weekend actually established', listing four tiers as cards. 'Confirmed by the files': 120 shards, 645 GB, bfloat16, MIT licence, 45 layers, 1,048,576-token ceiling. 'Confirmed by the vendor guide': 320B total, 18B active, GLM-5.3-Flash architecture, 8xH200 reference. 'Vendor-reported, unreproduced': 35 benchmark rows, harness undocumented. 'Not confirmed': no price, no hosted endpoint, no independent score. Footer: 'Sage-S2.1-Flash figures unaudited; file and licence facts from the Hugging Face repository.' The OrcaRouter logo is composited in the bottom-right corner.

What is confirmed, and what is not

Sorting the weekend's evidence into two piles is the most useful thing a reader can do with it.

• Confirmed by files — the weights exist and are downloadable under MIT; 120 shards, about 645 GB, bfloat16; 45 layers with a 34/11 linear-to-sparse attention split; 288 routed experts, top-8; a 1,048,576-token positional ceiling; a vision tower and image/video tokens; an FP8 sibling published two days earlier.

• Confirmed by the vendor's own guide — 320B total, 18B active; the GLM-5.3-Flash architecture; 8×H200 as the reference deployment; five documented serving traps.

• Vendor-reported, unreproduced — every benchmark score, including the 24-point BabyVision margin and the 72.2 on OfficeQA Pro V2.

• Not confirmed — what it costs. There is no price anywhere: not on the Hub, not in the guide, not on the microsite. The "Try the API" button points at discovery.intern-ai.org.cn, which returned a near-empty shell to a non-JavaScript fetch; we could not establish from outside whether the model is callable there, under what identifier, or at what rate.

• Also unconfirmed — whether this is a launch at all, in the sense the word usually carries. No announcement, no paper under this name, no partner access, and a model card that a script could have generated.

How to look at it without betting a production path on it

The practical question for most teams is not "is this better than what I run" but "is it worth an hour". At 645 GB in bfloat16 it is not a laptop download; the honest entry point is a rented 8×H200 node, an afternoon with the deployment guide, and the acceptance that you are running a model nobody has published a harness for.

The cheaper experiment, before any of that, is to read the architecture line as a shortcut. InternLM says in its own guide that the model is built on the GLM-5.3-Flash architecture and that you should follow Z.ai's serving recipes for it, substituting one repository name for the other. That means the base architecture — same 320B/18B shape, same hybrid sparse-and-linear attention, same million-token ceiling — is available to test today through a hosted API rather than through eight GPUs you have to find first. If the base model already behaves on your scientific documents, the derivative is a refinement worth the download. If it does not, you have saved the weekend.

Where OrcaRouter does earn its place here is in the gap between "we want to try it" and "we can serve it". Automatic failover lets you put an experimental endpoint behind a policy without making it the only path: if the new model returns a malformed tool call — and the deployment guide tells you exactly how it will, if you skip the flag or the patch — the request falls through to a model you already trust, and you keep the trace. That is a better way to collect evidence than a one-off eval script, because it collects evidence on real traffic.

What would settle it

Three events, in ascending order of usefulness. A model card would tell us what the lab thinks the model is for, and would probably carry the same table that is currently stranded on a GitHub Pages host. An entry on an independent leaderboard would let someone subtract the vendor's margins from a common harness. And a price — any price, anywhere — would turn Sage-S2.1-Flash from a download into a decision.

Until one of those lands, the accurate description is the one the files support: a large, MIT-licensed, scientifically-tuned multimodal MoE, built on a public architecture, published without a word of explanation, whose own authors are still patching the serving stack. That is not nothing. It is also not yet a product, and the difference matters most to the people who would have to run it.