Intern-S2-397B vs Intern-S2 Mobius: The 397B Flagship and the 35B Reasoner

Intern-S2-397B and Intern-S2 Mobius are both InternLM models, both Apache-2.0, both called Intern-S2, and they could hardly be more different underneath the name. Intern-S2-397B is the flagship: a 403-billion-parameter mixture-of-experts model for scientific intelligence and long-horizon agents that InternLM released on September 13, 2026. Intern-S2 Mobius is a 35-billion-parameter reasoning experiment that arrived on July 29, 2026, built on the Mobius-v0 architecture with a globally shared memory and iterating reasoners, continual-pretrained from Qwen3.5-35B. The name collision is the entire reason this comparison needs to exist, because a search for "Intern-S2 review" will surface both models and the two are not substitutes for each other in either direction.

Two architectures, two jobs

Intern-S2-397B is a conventional transformer MoE, wide where it matters: 60 layers, 512 experts with 10 active per token, a 256K text reasoning context and a 64K multimodal one, and an input surface that takes text, image and time series. Its pre-training paradigm reads raw pages of scientific literature — figures, layout, symbolic notation and prose together — and its reinforcement learning spans more than twenty scientific domains, plus long-horizon agent training in sandboxed environments. It is a specialist that is also a general model: MMLU Pro 89.77, MMMU Pro 81.68, SWE-bench-Pro 68.54 on its own card, all InternLM-reported and unreproduced.

Intern-S2 Mobius is a different bet on a different question. Instead of binding knowledge storage and reasoning computation layer by layer, the Mobius-v0 architecture keeps a globally shared Memory of knowledge vectors and lets multiple Reasoners iterate against it, refining hidden states over high-density continuous representations. The pitch is inference efficiency: a nearly 4x end-to-end speedup over the Qwen3.5-35B baseline it descends from, with comparable or stronger reasoning scores. That is a throughput claim about a specific baseline — and a vendor claim, with no independent reproduction — but it is the reason the model exists.

• Architecture — Intern-S2-397B: standard transformer MoE, 60 layers, 512 experts / 10 active vs Intern-S2 Mobius: Mobius-v0, shared Memory + iterating Reasoners.

• Scale — Intern-S2-397B: 403B total, ~807 GB of weights in BF16 vs Intern-S2 Mobius: 35B total.

• Lineage — Intern-S2-397B: original scientific pre-training paradigm vs Intern-S2 Mobius: continual-pretrained from Qwen3.5-35B.

• Inputs — Intern-S2-397B: text, image, time series vs Intern-S2 Mobius: text, image.

• Context — Intern-S2-397B: 256K text / 64K multimodal vs Intern-S2 Mobius: no independently stated context figure yet.

• Speed claim — Intern-S2-397B: none advertised vs Intern-S2 Mobius: nearly 4x end-to-end reasoning speedup over the Qwen3.5-35B baseline, vendor-reported.

• Independent evidence — neither model has a third-party score.

Which number belongs to which model

The name collision becomes a practical hazard the moment you read a review. A benchmark claim about "Intern-S2" — MMLU Pro 89.77, HMMT-2026 93.56, Biology-Instructions 55.71 — is about the 397B flagship, because that is the model the launch table belongs to. A throughput claim about "Intern-S2" — the nearly 4x speedup — is about Mobius, because that is the model whose architecture is about inference efficiency. Check the parameter count before quoting either number. A review that says "Intern-S2 is 4x faster" and one that says "Intern-S2 beats Grok 4.6 on TerminalBench" are talking about different models, and the second claim does not even survive its own vendor table.

On the evidence side, neither model has been tested by anyone outside InternLM. The flagship's card is a vendor benchmark table run through OpenCompass, VLMEvalKit and AgentCompass; the Mobius card is a description plus a speedup claim against a baseline the model was fine-tuned from. The absence of independent verification matters more here than in most comparisons because the two models' strongest claims live on completely different axes — capability for one, throughput for the other — and neither axis has an outside measurement yet.

Which one to run

Run Intern-S2-397B when the job is scientific understanding: reading a figure-dense paper, analyzing a molecular diagram, forecasting from a time-series signal, or running a long agent that has to keep working through a research task. It is the model this week's coverage is about, and its general rows are respectable enough that you are not buying a one-trick pony. The price is hardware: a 403B checkpoint needs a serious multi-GPU node, and if you do not own one the official Intern API is the alternative.

Consider Intern-S2 Mobius when the job is high-volume reasoning on hardware you already own and you care about per-token cost more than a point or two of benchmark score. A 35B model with a plausible 4x throughput advantage over its baseline is a genuinely interesting serving proposition — but validate the speedup on your own traces before you build around it, because nobody outside InternLM has reproduced it yet. Either way, label the model explicitly in every write-up, because the name alone no longer says which one you mean.