Muse Wonder vs Intern S2: 1,048,576 Tokens Against 256K, and No Shared Ruler

Intern-S2-397B and Muse Wonder have never been measured against each other by anyone. That is not a gap in the research — it is a structural fact about how each model came to exist. I​nternLM published Intern-S2-397B on September 13, 2026 as an Apache-2.0 open-weights checkpoint for scientific intelligence and long-horizon agents, benchmarked entirely by its own makers, with no third-party score at all. Meta Superintelligence Labs began selling Muse Wonder on October 12, 2026 through the Meta Model API and Muse Code, and it has exactly one independent evaluator's row. So everything below requires labelling which side of the fence each number sits on, and the two models disagree about something more basic than benchmark scores: what a context window is for. Muse Wonder's 1,048,576 tokens are a flat, priced, single-response budget. Intern-S2-397B's are split — 256K for text reasoning and 64K for multimodal input — on a checkpoint you download rather than call.

The useful version of this comparison is not which model wins. It is that one of them is an instrument for scientific document work that you can run inside your own perimeter, and the other is a rented generalist with a very large flat-rate window. Those are different purchases.

What each model actually is

Intern-S2-397B is Shanghai AI Laboratory's I​nternLM flagship, released September 13, 2026 under Apache-2.0. The name carries 397B; the model card says 403 billion total parameters, and that difference is the kind of inconsistency a checkpoint built in public accumulates. Architecturally it is a standard mixture-of-experts transformer — 60 layers, 512 experts with 10 active per token — and its headline claim is a pre-training paradigm that reads raw pages of scientific literature, figures, layout, symbolic notation and prose together, rather than parsing the text out first. Around that it scales reinforcement learning across more than twenty scientific domains and adds long-horizon agent training in sandboxed environments. It takes text, image and time-series input, produces forecasts from the time-series path, serves on LMDeploy, vLLM and SGLang, supports tool calling, and thinks by default unless you pass enable_thinking=False per request. In BF16 the checkpoint is roughly 807 GB, so self-hosting means a serious multi-GPU node.

Muse Wonder is Meta Superintelligence Labs' frontier release of October 12, 2026, announced as a two-checkpoint series. Only muse-wonder-1.0 has a per-token endpoint, on the Meta Model API and inside Muse Code; Muse Wonder Full, the higher-scoring configuration, is sold only through a Muse Code subscription and has no public rate card. The callable checkpoint takes text, image, speech and video input, returns text, carries a 1,048,576-token context window, caps a single response at 943,718 tokens, prices at $4.00 per million input tokens, $0.20 cached and $20.00 output, and runs reasoning mandatorily on an effort ladder from minimal to max that defaults to medium. There are no weights, no licence, and no version you can keep running if the endpoint changes.

That is the first real difference between them, and it is not a benchmark row. One is a file. The other is a contract.

Context window, which the two models mean differently

This is where a spec sheet comparison misleads most, because both models are routinely described as long-context.

• Text context — Muse Wonder 1,048,576 tokens in one window; Intern-S2-397B 256K for text reasoning

• Multimodal context — Muse Wonder's full window covers its image, speech and video input; Intern-S2-397B carries a separate 64K multimodal context

• Maximum output in one response — Muse Wonder 943,718 tokens; Intern-S2-397B not published as a comparable figure

• Input surface — Muse Wonder text, image, speech, video; Intern-S2-397B text, image, time series

• Time-series forecasting — Muse Wonder none; Intern-S2-397B supported, currently wired through LMDeploy only

• Rate — Muse Wonder $4.00 in, $0.20 cached, $20.00 out per million tokens; Intern-S2-397B no published per-token rate, because the distribution model is a download

The forecasting row is the one that has no counterpart on the other side, and it is not a marketing feature — it is the reason the model exists in the form it does. A model that ingests a time-series signal and produces a forecast is doing something a text model cannot be asked to do, whatever its window size. If that is your workload, the comparison is already over.

For everyone else the relevant question is what actually fits. A 1,048,576-token window at a flat rate is a different product from a 256K text window on hardware you own, and the difference shows up in exactly the workloads that made long context interesting — a codebase in the prompt, a corpus of papers, a long agentic session that keeps its history.

The evidence bases, which do not touch

The two models' published numbers cannot be placed on the same axis, and it is worth being blunt about why.

Intern-S2-397B's benchmark card is I​nternLM's own, run through OpenCompass, VLMEvalKit and AgentCompass by the team that built the model, with no independent reproduction. Its columns are Qwen3.5-397B-A17B, D​eepSeek V4 Pro, Kimi K2.7-Code, GLM-5.2, GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.8 — no model from M​eta, and no Muse Wonder, so the vendor's own table does not contain a single Muse comparison. and any "Intern-S2 versus Muse Wonder" figure circulating anywhere is an extrapolation across two unrelated tables.

What that card does publish, all of it I​nternLM-reported: 89.77 on MMLU Pro, 93.56 on HMMT-2026, 81.68 on MMMU Pro, 68.54 on SWE-bench-Pro, 64.04 on TerminalBench 2.1, and on the scientific sets 55.71 on Biology-Instructions, 63.28 on SciReasoner 1.5 and 53.95 on Mol-Instructions. Read that as a portrait: strong on knowledge, maths and multimodal reasoning, leading on software engineering, and twenty points behind Claude Opus 4.8 on terminal work — the exact axis a production agent spends its day on.

Muse Wonder's evidence is the inverse shape. M​eta's launch table covers both checkpoints at maximum reasoning and is built largely from sustained-agent benchmarks — DeepSWE 1.1, SWE-Marathon, Terminal-Bench 4.0-AA, OSWorld V2.1 — all of it vendor-reported, in M​eta's harness, unreproduced. Independent commentary has repeated the strongest rows without the caveat, which is how a vendor harness becomes a fact in public. The one genuinely independent row is Artificial Analysis' Intelligence Index at 52 at maximum effort, ninth of more than 200 models, with about $4.52 of measured cost per completed index task and roughly 130 million output tokens consumed against a board median near 85 million.

So: one model's strength is asserted by its maker on scientific benchmarks no third party has rerun. The other's, on agentic benchmarks, the same. Neither has been compared to the other by anyone. The only thing an honest page can do with that is say it out loud.

Where the comparison is decided by price, and where it is not

Intern-S2-397B has no rate card. That is the whole pricing section, and it is a real finding rather than a missing value: you either self-host roughly 807 GB of BF16 weights on hardware you own, or you register for a token on I​nternLM's own API and live within whatever quota is granted. There is no per-token number to put beside Muse Wonder's $4.00 and $20.00, and no invoice to compare with a pay-per-token bill.

What that does to the decision is subtler than "free versus paid." A downloaded checkpoint converts a variable cost into a capital cost, and it converts a vendor relationship into an artefact you own. That is worth a great deal when your data cannot leave your perimeter, when your workload is steady enough to justify the GPUs, or when reproducibility matters more than a benchmark point. It is worth very little if your traffic is spiky, if your team has no serving stack, or if you need the model to improve without you doing anything.

Neither model is served by us, and it is worth saying so plainly rather than implying otherwise: Intern-S2-397B is a brand-new Apache-2.0 checkpoint with no routing coverage, and Muse Wonder is on M​eta's API and inside Muse Code with no third-party serving we can verify. What a router does help with is the half of this that is callable. OrcaRouter carries more than 200 models behind one key at provider list price with 0% markup — no middleman delta, so a vendor price cut is live the same day — including M​eta's own Muse Spark 1.2 and Muse Spark 1.1, so the M​eta models you can reach today sit next to everything else on the same credential. If your production traffic lands on a routed model and your scientific batch work runs on Intern-S2-397B somewhere you control, that seam is a routing decision rather than a second integration.

Who should pick which

Take Intern-S2-397B when the work is reading science: a figure-dense paper, a molecular diagram, a scanned table, a forecast from a time-series signal, or an agent that has to keep working through a research task. The Apache-2.0 licence means you can fine-tune it on proprietary data and keep the result inside your own perimeter, which is the part of this comparison no rate card captures. Budget for the hardware and expect no independent scoreboard for some months — the repository is weeks old and carries no third-party measurement.

Take Muse Wonder when the work is general, long and agentic, and when you would rather rent than own: a 1,048,576-token window at a flat rate with no step, a 943,718-token response ceiling that nothing in Intern-S2's published material answers, mandatory reasoning for hard multi-step tasks, and a pay-per-token bill you can put in a spreadsheet today. The caveat is the same one that applies to every row M​eta published: two days of independent measurement is two days, and one evaluator is one evaluator.

The team that needs both should treat the boundary between them as a routing question rather than a migration. Run the general traffic on a model you can call, keep the scientific work on the checkpoint you can hold, and re-run your own evals on Intern-S2-397B before believing a vendor figure about its terminal performance — because the vendor's own table is the least flattering row in it.