
Sage-S2.1-Flash vs Qwen 3.8 Flash: Eight H200s Against Fifteen Cents
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 152 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 119 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
The model card for Sage-S2.1-Flash is twenty-one bytes long. It opens with a licence header, says license: mit, and stops. There is no introduction, no benchmark table, no context length, no intended use, no contact address. The repository itself is not empty — 120 shards of bfloat16 weights, about 645 GB of them, plus a configuration file, a chat template, a deployment guide and five small scientific adapter folders — but everything a reader would normally use to decide whether to care has been left out. The weights went up on 15 November 2026 and were revised through the 16th. InternLM, the Shanghai AI Laboratory's model organisation, has announced nothing: no blog post, no paper, no post on the family's own site. At the time of writing the repo shows zero downloads for the month, one like, and no inference provider serving it.
Qwen 3.8 Flash is the opposite kind of object. Alibaba has sold it through QwenCloud since 26 August 2026 at $0.15 per million input tokens and $0.47 per million output, published its rate limits and its tool list, shipped open weights under the Qwen3.8-Flash-Next name, and put a twelve-row benchmark table on that card. It is live on OrcaRouter as qwen/qwen3.8-flash at the vendor's own list price. You can send it a request in the next thirty seconds.
So this is not a comparison of two models you can buy. It is a comparison between a cheap hosted model with a published price and a 321-billion-parameter open-weights mixture of experts with no price, no score and no announcement — and the honest question is not which one is better, because nobody can answer that yet. It is whether standing up eight H200s for the unmeasured one is a better use of money than paying fifteen cents per million tokens for the measured one, and for how many teams that answer is ever yes.
What the Sage-S2.1-Flash repository actually proves
Facts first, because they are more plentiful than the empty card suggests. The parameter count in the weight index is 321,323,031,390, and the deployment guide rounds that to "320B-A18B" — 320 billion total, 18 billion active per token. The text configuration carries 45 layers, of which 34 are linear-attention layers of the gated delta-net kind and 11 are sparse-attention layers in the DeepSeek lineage, with 288 routed experts, eight of them plus one shared expert firing per token, a hidden size of 4,096 and a 154,880-token vocabulary. Maximum position embeddings are set to 1,048,576.

The licence is MIT, with no acceptable-use rider in the repository. That is worth pausing on: a 321B multimodal MoE under a permissive licence, with no gate on the download, is a rarer object than the silence around it suggests.
The lineage is checkable and slightly surprising. The architecture string is glm5_next, the class is Glm5NextForConditionalGeneration, and the deployment guide says outright that the model uses the GLM-5.3-Flash architecture and that the framework recipes to follow are the official GLM-5.3-Flash ones, with the model path swapped. The older, larger Sage-S2.1 — 753B parameters, architecture glm_moe_dsa, still carrying an "Atria Dawn" model card that describes it as built on a 744B GLM-5.2 backbone — is a different generation of backbone entirely. Flash is not a distilled S2.1. It is a new, smaller model built on a newer architecture from a different lab, wearing the same family name.
That family name is itself a moving target. The old internlm/Atria-Dawn repository now redirects to internlm/Sage-S2.1, and the redirect target's card still says "Atria Dawn" in every heading — the rename happened without the documentation following it. Anyone who read the Atria Dawn coverage earlier in the year and is now looking for "Sage" should understand those are the same repository under two names, and that Flash is its new, leaner sibling.
The scientific adapters are the part of the repository that no vendor announcement would have explained anyway. Under projectors/sage-s2.1-flash/ sit five small folders — cif/jarvis, dna/splice, ecg/phu01, eeg/neu06, protein-seq/mpbind — each holding a configuration file and a projector of four to twenty-eight megabytes that maps a scientific feature stream into the backbone's 4,096-wide space. Crystal structures go through a CHGNet 0.3.0 encoder, one 64-dimensional state per atom. Protein sequences go through ESM2-t33-650M, one last-layer state per residue. The DNA adapter is a task-adapted DNABERT2 readout over three features; the ECG and EEG adapters are similarly small and task-specific. Total weight cost for all five: under 90 MB, against 645 GB for the language model.
Those configuration files also leak the provenance the model card withheld. They name the backbone checkpoint directory, the training-run date stamps, and a fingerprint reading sage-s2.1-flash-hf111-bf16-norm-261113 — 13 November 2026, two days before the weights appeared. That is how the lineage is established here: not from a press release, but from a config field the packaging forgot to strip.
What Qwen 3.8 Flash publishes, and what it does not
The contrast is not that Qwen is honest and Sage is not. It is that one vendor documents its product and the other has not got around to it. Alibaba's card for Qwen 3.8 Flash specifies a 1,000,000-token context with a 991,000-token maximum input, a 131,072-token maximum output, and a 262,144-token reasoning budget; multimodal input across text, image and video with text output; function calling, structured JSON output, prefix completion, context caching, batch processing and fine-tuning; and both OpenAI and Anthropic request protocols, which is why it drops into Claude Code and Codex unchanged.
The open-weights artefact behind it is Qwen3.8-Flash-Next, published 24 August 2026: 125B parameters with 6 billion activated, plus a 51B n-gram embedding table and a 4B multi-token-prediction head, 48 layers arranged as twelve repeats of three gated-delta-net blocks and one sparse-attention block, 512 experts with ten routed plus one shared active per token, and 262,144 tokens of native context extensible to a million. Alibaba describes the hosted Qwen 3.8 Flash as that architecture plus production features — the default 1M window and the built-in tools are the additions. The weights sit under the Qwen community licence, not MIT.
The benchmark table on that card is vendor-reported and has not been independently reproduced. Read it with that label attached: DeepSWE 1.1 at 58.7, SWE-bench Pro at 62.5, SWE-bench Multilingual at 81.0, NL2Repo-Bench at 48.1, CoWorkBench at 73.9, JobBench at 55.7, Toolathlon Verified at 73.5, IFBench at 81.3, GPQA Diamond at 91.7, HLE at 35.9, LiveCodeBench v6 at 91.9, and Agents' Last Exam at 24.3 pass@1 with a 51.2 score. Those numbers are Alibaba marking its own homework across a comparison set that includes a 397B predecessor and a Claude model. They are still a great deal more than Sage-S2.1-Flash offers, which is nothing at all.

Set the two side by side on the dimensions a purchase decision actually turns on:
• Total parameters — Sage-S2.1-Flash 321B, of which 18B active per token, versus Qwen 3.8 Flash 125B plus a 51B n-gram table, of which 6B active
• Context — 1,048,576 tokens configured versus 1,000,000 published, with a 991,000 maximum input and 131,072 maximum output on the Qwen side
• Input price — Sage-S2.1-Flash has none, because nothing serves it, versus $0.15 per million tokens
• Output price — none versus $0.47 per million, with cached reads at $0.0184 per million on our listing
• Licence — MIT, ungated, versus the Qwen community licence
• Published evaluation — zero benchmarks versus a twelve-row vendor table, unreproduced
• Serving weight — 645 GB bfloat16 across 120 shards, or 328 GB in the FP8 sibling, versus a hosted endpoint
The price comparison that is really a capital-expenditure question
There is no per-token price for Sage-S2.1-Flash to compare against $0.15, because the model has no hosted endpoint anywhere. What it has is a serving requirement: eight H200s for the bfloat16 checkpoint, per the deployment guide's own reference configuration. So the comparison is not cents against cents, it is rent against tokens, and it resolves to a single number nobody has published.
Take a representative $3 per GPU-hour for an H200 — an assumption, not a quote, and the figure moves by cloud and by commitment. Eight of them is $24 an hour, $576 a day, about $17,500 a month before you have served a single request. At Alibaba's rates, $576 buys 3.84 billion input tokens or 1.23 billion output tokens. So the break-even question is whether your traffic would otherwise spend more than $576 a day on Qwen 3.8 Flash — roughly 3.8 billion input tokens' worth, every day, sustained.
That comparison is only valid if you know how many tokens per second Sage-S2.1-Flash produces on eight H200s. Nobody has measured it and the repository does not say. Active parameter count gives you a rough intuition — 18B active against 6B active suggests the Sage model costs about three times more compute per generated token — but intuition is not a throughput figure, and without one the break-even arithmetic above has an unknown on both sides. The FP8 sibling cuts the resident weight from 645 GB to 328 GB, which changes what hardware you need; it does not tell you how fast the result runs.
For most teams that unknown is the whole argument. A model you cannot benchmark for throughput is a model you cannot cost, and a model you cannot cost is not a candidate for a production path — it is a research object. That is not a criticism of the release; it is what an unannounced drop is for.
The evidence asymmetry, stated plainly
Two models, two failure modes of evidence. Qwen 3.8 Flash has vendor-reported benchmarks and eighty-two days of third-party use, which means its numbers are unaudited but its behaviour is at least observable: there are people running it, reporting on it, and hitting its limits in public. Sage-S2.1-Flash has neither. Its 645 GB of weights are more evidence of a real model than a leak or a screenshot would be — this is a shipped artefact, not a rumour — but weights are evidence of existence, not of quality.
The specific things that remain unknowable today: what Sage-S2.1-Flash scores on anything; how fast it decodes; whether its 1M position embeddings hold up at length or degrade the way most long-context extensions do; whether the vision tower in its configuration is trained to production standard or carried over from the backbone; whether the scientific projectors are evaluated on anything beyond their own training distributions; and what InternLM intends to charge, if anything, for a hosted version. The repository answers none of these. Anyone who tells you otherwise is extrapolating from the backbone's reputation, which is a reasonable prior and not a measurement.
What the model is actually for
The projectors are the tell. Nobody bolts five scientific modality adapters onto a general chat model as an afterthought; the folder structure, the per-task prompt templates and the encoder choices describe a model built to read crystals, splice sites, cardiac traces, brain traces and protein sequences as first-class inputs rather than as text descriptions of those things. Crystal structures are tokenised as atom states with a 512-atom ceiling; protein sequences carry per-residue states with an optional scientific readout head; the DNA adapter is a probability readout over a single object.
Qwen 3.8 Flash does not do this. It reads images, video and text, and it does so with published numbers on visual benchmarks — CharXiv, MathVision, ERQA, LVBench, OSWorld 2.0 — but a crystal lattice rendered as a PNG is not a crystal lattice. If your workload is structural biology, materials screening or physiological signal analysis, the honest reading is that these two models are not competing for the same job and the Qwen comparison is a price anchor rather than a substitute. If your workload is text and code, the projectors are irrelevant and Sage-S2.1-Flash is a 321B chat model with no card.
One caution on the multimodal question, because the configuration invites over-reading. The repo carries a 24-layer vision tower at hidden size 1,024 with 448-pixel inputs and 14-pixel patches, image and video start and end tokens, and a chat template with image, video and audio slots. But the template is the GLM-5.3 family template, the audio slots have no corresponding audio configuration, and the deployment guide's only media-specific instruction is about images arriving inside tool results. Treat the tower as present and undocumented: weights are built to take more than text, and what the release intends you to feed it is unstated.
The friction, which is not zero
Self-hosting Sage-S2.1-Flash is not a matter of pointing vLLM at a repository. Three requirements come from the deployment guide, and the second one is the awkward one.
• --trust-remote-code is mandatory on every engine — the repository ships its own tokenizer and processor, and without the flag the engines silently load a stock tokenizer that encodes strings like <tool_call> as control tokens rather than text, breaking reasoning and tool calls with no warning
• vLLM 0.33.0 needs a supplied patch to vllm/renderers/hf.py, because when parallel tool results come back out of call order the template renders them in call order while vLLM collects their images in message order, silently attaching each image to the wrong result; the upstream fix is still a draft pull request
• LMDeploy needs transformers between 5.17 and 5.19, because 5.19 renames the layer type the engine will not accept; SGLang needs 0.5.20 or later plus the repository's chat template passed explicitly, because SGLang otherwise installs its own built-in template for the architecture
Every one of those is a documented workaround for a model that is a day old, and the guide is candid that the frameworks are still moving. None of it is unreasonable; all of it is engineering time that Qwen 3.8 Flash does not ask for, because on the hosted side it is a base URL and an API key.
One key, two very different levels of commitment
Qwen 3.8 Flash is on OrcaRouter as qwen/qwen3.8-flash at Alibaba's list price passed through with no markup — $0.15 per million input tokens, $0.47 per million output, $0.0184 per million for cached reads, a 1,000,000-token context and a 131,072-token output ceiling. QwenCloud's own page quotes $0.016 for the implicit cache variant; the pass-through rule means whatever Alibaba changes, you get on the same day rather than at the next contract renewal. Sage-S2.1-Flash is not in the catalogue and nothing serves it, so there is no route to it here and no reason to pretend otherwise.

What the routing layer is genuinely good for in this situation is keeping the evaluation cheap and reversible. Run the Sage experiment on your own hardware if you have it, keep Qwen 3.8 Flash as the production route, and let a failover chain sit behind the new path so that a model with an empty card never becomes a single point of failure on the first day it handles real traffic. The same idea applies to the harder question — model fusion, where a panel answers one prompt with several models and a judge reconciles them, is the cheapest way to find out whether a 321B scientific MoE is worth its rent: you do not migrate, you ask both, and you read the answers.
Who should pick which
If you are shipping text or code this quarter, Qwen 3.8 Flash is the defensible answer and Sage-S2.1-Flash is not an answer at all. The Qwen model is cheaper per token than almost anything at its capability level, documented well enough to plan against, licensed clearly enough to deploy, and observable in public. Its benchmarks are Alibaba's own and should be treated as claims, but a claim plus eighty-two days of other people's experience beats a silence plus three days of nobody's.
If you work on structural biology, protein design, materials or physiological signals, the calculus inverts. The five projectors in the Sage-S2.1-Flash repository are a capability neither Qwen 3.8 Flash nor most open models offer at any price, the licence is MIT, and the hardware requirement — eight H200s, or fewer with the FP8 checkpoint — is the kind of cost a research group already carries. The right move is to download it, run the CPU-side projector checks the repository ships, and measure the two things it does not tell you: decode throughput on your hardware, and accuracy on your own held-out set. That is a week of work and it is the only way this model becomes a decision rather than a curiosity.
The third case is the largest one. If you are merely curious, the correct posture is to wait for a model card. When InternLM publishes one, the numbers that matter will be the ones this repository cannot supply — throughput, long-context degradation, whether the vision tower is live, and what the projectors score outside their training distributions. Until then Sage-S2.1-Flash is a real artefact with an unread specification, and the most useful thing about it is that the weights exist, are ungated, and are MIT — which means that when somebody does measure it, you will not need permission to check their work.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
