A generated title card reading 'Sage-S2.1-Flash vs DeepSeek V4.1 Flash' above the subtitle 'Two open-weight MoEs, one vendor scoreboard', with two chips below reading '18B active' and '16B active'. 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 vs DeepSeek V4.1 Flash: The Only Matchup With a Shared Scoreboard

Author

jinhao.song

Date Published

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

There is exactly one place on the internet where Sage-S2.1-Flash and DeepSeek V4.1 Flash have been scored against each other, and it is not a neutral one. InternLM, the lab that published Sage-S2.1-Flash on 15 November 2026, put DeepSeek-V4.1-Flash-1017 into its own comparison table — 35 benchmarks, eight rival models, a tab-separated file in a GitHub repository that the Hugging Face model card does not link to. So the two models share a scoreboard, and one of them built it.

That is more than can be said for either of the other two comparisons in this series, where no shared table exists at all. It also makes this the most useful of the three pages, because a common table — even a vendor's own — lets you ask questions that a pair of separate blog posts cannot answer. What follows is what the shared table says, what the two architectures underneath it are, and how much of it you should be willing to believe.

The asymmetry that frames everything

Put the two releases side by side as artefacts before you put them side by side as models.

DeepSeek V4.1 Flash arrived on 10 September 2026 with a full model card, a technical report PDF titled "Pushing the Limits of KV Cache Compression", an MIT licence, a documented API with published off-peak and peak rates, and a refreshed snapshot — 1017 — on 17 October. Its Hugging Face repository has been downloaded 1.69 million times and carries 4,978 likes. Artificial Analysis scores it independently at 32.6 on the Intelligence Index at maximum effort.

Sage-S2.1-Flash arrived on 15 November 2026 with 120 bfloat16 shards, a 21-byte model card whose entire contents are a licence declaration, a deployment guide, and no published price anywhere. It has one like and zero downloads. No independent harness has touched it.

Both are MIT-licensed open-weights mixtures of experts with a one-million-token context window and native image input. On paper they are peers. In evidence they are not close.

A generated two-column comparison scoreboard titled 'Sage-S2.1-Flash vs DeepSeek V4.1 Flash - the scoreboard'. The Sage-S2.1-Flash column reads: Parameters: 320B, 18B active; Layers: 45, 34 linear; Price: none published; Context: 1,048,576 tokens; Independent score: none; Downloadable: MIT weights. The DeepSeek V4.1 Flash column reads: Parameters: 552B, 16B active; Layers: 40, encoder-decoder; Price: $0.15 / $0.60 off-peak; Context: 1,000,000 tokens; Independent score: AA Index 32.6; Downloadable: open weights. Footer: 'Sage rows from InternLM; DeepSeek price from DeepSeek; DeepSeek score per Artificial Analysis.' The OrcaRouter logo is composited in the bottom-right corner.

What the shared table actually contains

InternLM's benchmark file lists 31 rows that carry a DeepSeek-V4.1-Flash-1017 figure. Reading it is a different experience from reading a launch post, because the rows where Sage loses are still there. A representative slice, Sage first in every line:

• Terminal-Bench 2.1 — 87.9 vs 87.6, a three-tenths tie

• DeepSWE — 67.0 vs 73.9, a 6.9-point deficit

• SWE-bench Pro — 60.6 vs 64.7, a 4.1-point deficit

• NL2Repo-Bench — 57.2 vs 62.7, a 5.5-point deficit

• APEX-Agents — 61.7 vs 69.2, a 7.5-point deficit

• WideSearch — 81.5 vs 86.3

• BrowseComp — 91.2 vs 93.8

• CyberGym — 84.3 vs 88.8

• τ³-bench Banking — 45.4 vs 47.4

• OSWorld v2.1 — 69.3 vs 70.1

• AutomationBench — 56.7 vs 57.8

The pattern in that list is uncomfortable for the newer model: on general agentic and coding work, DeepSeek V4.1 Flash wins the majority of rows, often by margins that are not noise. A model published two months earlier, by a lab with a public technical report, beats the quiet release on the benchmarks the quiet release itself chose to publish.

Then the table inverts. The rows where Sage leads are the scientific and document ones, and the margins are of a completely different order:

• Harvey LAB — 25.0 vs 11.7, more than double

• OfficeQA Pro V2 — 72.2 vs 69.7

• BabyVision — 75.8 vs 58.0, a 17.8-point gap

• OSWorld-G — 84.8 vs 61.4, a 23.4-point gap

• MLE-bench Lite — 90.3 vs 89.2

• Workspace-Bench-Lite — 67.1 vs 66.8

• GDPval — 78.0 vs 77.6

• SpreadsheetBench 2 — 37.1 vs 35.8

Two patterns, and they point in opposite directions. On software engineering and tool-driven agent work in a general environment, DeepSeek V4.1 Flash is the stronger model and InternLM's own numbers say so. On scientific perception, document-heavy professional work and graphical desktop control, Sage-S2.1-Flash is not marginally ahead but in a different band.

Why the architectures would produce exactly that split

The two models are solving different problems, and their configs show it before any benchmark does.

DeepSeek V4.1 Flash is a 552-billion-parameter backbone with 16 billion active per token, built on what DeepSeek calls a Causal Encoder-Decoder architecture: 40 layers split evenly into a 20-layer causal encoder and a 20-layer decoder. The technical report's framing is KV cache compression — the model's headline claim is about serving efficiency at a million tokens, not about perception. It processes images natively, but its published emphasis is long-context serving economics.

Sage-S2.1-Flash is a 320-billion-parameter model with 18 billion active per token and, per its deployment guide, 45 layers. The config divides them sharply: 34 layers use a linear-attention type the file calls kda, and only 11 use sparse attention with a 2,048-token top-k index. There is a 24-layer vision tower at 448-pixel input, and the token vocabulary carries explicit image, video, audio, start and end tokens. InternLM's own English release copy — hosted at internlm.github.io/Sage-S2.1-Flash, created on 16 November, unreferenced by the model card — describes training across molecular graphs, crystal structures, protein sequences and 3D structures, DNA and RNA, spectra, mass spectra, X-ray diffraction, electrocardiograms, electroencephalograms and light curves.

Read that list against the rows Sage wins and the picture stops looking like benchmark luck. Harvey LAB is legal work over documents. OfficeQA Pro is financial-document question answering. BabyVision is visual perception. OSWorld-G is graphical desktop control — clicking a UI, not calling an API. Those are the constructs a model trained on scientific figures, spectra, plots and software interfaces would be expected to be good at, and they are precisely where a model trained for KV-cache-efficient general serving would be expected to lag.

The point cuts both ways. Sage loses the coding rows because it is not primarily a coding model; DeepSeek loses the vision rows because it is not primarily a perception model. Neither result is a scandal. The mistake would be reading either column as a general verdict.

Screenshot of the OrcaRouter model page for deepseek/deepseek-v4.1-flash. It shows a 1M-token context window, text and image input, vision, tools, JSON and reasoning tags, a release date of 2026-10-17, and the price row $0.15 input and $0.60 output per million tokens with a 1.20 s time to first token. The page states the model carries a context window of 1,048,576 tokens, can generate up to 384,000 tokens in a response, and is billed at $0.15 per 1M input and $0.60 per 1M output tokens.

The price spread, which is not close

DeepSeek V4.1 Flash has a published rate, and it is one of the cheapest rates on any model of this size:

• Off-peak, cache miss — $0.15 per million input, $0.60 per million output

• Peak, cache miss — $0.30 per million input, $1.20 per million output

• Cache hit — $0.003 per million off-peak, $0.006 peak

• Peak hours — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays

• Context — 1,000,000 tokens, with up to 384,000 output tokens

Sage-S2.1-Flash has no published rate. There is no price on the model card, none in the deployment guide, and none on the release microsite; the site's "Try the API" link points at a portal that would not render its catalogue to a non-JavaScript client. What Sage has instead is a hardware requirement: the guide's reference deployment is eight H200 GPUs at bfloat16, roughly 645 GB of weights across the node.

That makes the cost comparison a shape rather than a number. DeepSeek V4.1 Flash is $0.15 per million input tokens at off-peak, which means a million-token prompt costs fifteen cents. Running Sage-S2.1-Flash at a million tokens costs whatever eight H200s cost you for the wall-clock time the call takes, and nobody has published a throughput figure for it. If you already own the node, the marginal cost of the call is near zero and DeepSeek looks expensive by comparison. If you do not, DeepSeek wins by a margin so wide that the Sage comparison is not worth doing until a hosted price exists.

What the shared table cannot tell you

One property of a vendor-authored comparison is worth stating clearly, because it affects how to read every row above: the harness is undocumented. InternLM published the numbers, not the run configuration. There is no statement of how many attempts per task, what scaffolding was used, which sampling parameters, whether tools were available, or what the reasoning budget was for either model. DeepSeek's own 32.6 on the Artificial Analysis Intelligence Index comes from a public harness that applies the same procedure to every model in its index; InternLM's 73.9 for DeepSeek on DeepSWE comes from a private one.

That does not make the table useless. It makes it directional. The coding rows and the vision rows are both large enough, and internally consistent enough, that scaffolding choices probably are not generating them. The three-tenths tie on Terminal-Bench 2.1, on the other hand, is well inside the range any harness variation could produce — treat a 0.3-point difference as a tie and nothing is lost.

Also worth noting: InternLM published a figure for DeepSeek on DeepSWE (73.9) that is high relative to the other coding rows in its own table, where DeepSeek lands at 64.7 on SWE-bench Pro. That internal inconsistency is the sort of thing a published harness would resolve in an afternoon and a private one cannot resolve at all.

Screenshot of the Hugging Face model page for internlm/Sage-S2.1-Flash. The header shows the safetensors and glm5_next tags, an MIT licence, a model size of 321B params and a bfloat16 tensor type. The sidebar notes that the README exists but its content is empty, that the model is not deployed by any Inference Provider, and it lists the collection including internlm/Sage-S2.1-Flash. Download counts are not shown.

Deployment: one afternoon versus one document

If you are choosing on operational grounds rather than benchmark rows, the gap is wider than the scores suggest, and it runs the same direction as the evidence gap.

DeepSeek V4.1 Flash ships with a technical report, a standard transformers architecture, an fp8 variant, and deployment tags for Azure and SageMaker on its model card. It has been in the wild for two months, has 1.69 million downloads, and every serving quirk it has, someone has already hit and written down.

Sage-S2.1-Flash ships with a deployment guide that is, read carefully, a list of silent failures. Omit --trust-remote-code and the stock tokenizer encodes agentic strings like <|user|> and <tool_call> as control tokens — the model stops early or corrupts a tool call, and nothing is logged. Serve it on vLLM without the one-file patch in vllm_patch/ and parallel tool results returning out of order attach images to the wrong result, so the model answers about the wrong picture; the upstream fix is pending as vllm-project/vllm#67108. Serve it on SGLang without passing chat_template.jinja explicitly and the engine installs its own template for the architecture. On Hopper, FP8 KV cache is unsupported. LMDeploy needs transformers pinned between 5.17 and 5.19, because 5.19 renamed a layer type.

All of that is documented, which is to the lab's credit — it is more than most quiet releases manage. But it is also the fingerprint of a model that its own authors are still the only people to have served.

The one useful thing the shared table settles

Given all of the above, here is the decision the evidence actually supports.

• If your work is code, agent tooling, or general long-context serving — DeepSeek V4.1 Flash. It wins most of those rows in the rival's own table, it costs $0.15 per million input tokens off-peak, and it has an independent score to check against. There is no serious case for Sage here.

• If your work is scientific figures, spectra, plots, scanned professional documents or GUI-driven software — Sage-S2.1-Flash is worth the weekend. The 23-point OSWorld-G gap and the 13-point Harvey LAB gap are the kind of difference that changes whether a task is automatable at all, and no other model in InternLM's table comes close on those rows.

• If you need to decide today — DeepSeek V4.1 Flash, because it is the only one of the two you can call without a maintenance window.

DeepSeek V4.1 Flash is routed on OrcaRouter as deepseek/deepseek-v4.1-flash at $0.15 in and $0.60 out per million tokens — DeepSeek's own off-peak list price passed through with no markup — with a 1,048,576-token context, alongside the 1017 snapshot at the same rate. Sage-S2.1-Flash is not in our catalogue; it is open weights you would serve yourself, and we are not going to imply otherwise.

What a router buys you here is narrower than a migration and more useful than a benchmark. The split this article describes — general agentic and coding work to one model, scientific and GUI-heavy work to the other — is exactly the shape a routing policy expresses: one endpoint, a rule that sends image- and document-bearing requests down the specialised path, and automatic failover that keeps a provider-side degradation from taking the whole path down. Both DeepSeek identifiers sit on the same key, so trying the 1017 snapshot against the original costs a model-string change rather than a second contract. The piece you cannot route — Sage-S2.1-Flash — is the piece you would have to stand up yourself, which is the clearest argument for prototyping against what is on the catalogue first.

What would make this page worth rewriting

Two things. A price for Sage-S2.1-Flash, which would turn the cost section from a shape into arithmetic. And an independent score, ideally on the same harness that put DeepSeek V4.1 Flash at 32.6 — because the interesting question is not whether Sage beats DeepSeek on average. Everything above suggests it does not. The interesting question is whether Sage's 23-point OSWorld-G advantage survives a harness that was not written by the lab that published it, and whether the 6.9-point DeepSWE deficit is as real as the coding rows in its own table imply.

Until someone reruns those rows, the shared scoreboard is the best evidence available and it was built by one of the two competitors. Read it the way you would read any self-published comparison that happens to include its own losses: more trustworthy than a table with no losses in it, and still not the same thing as a measurement.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily