A title card reading NVIDIA Nemotron 3.5 Super VL 120B vs Intern-S2-397B, with four statistic panels giving Intern-S2-397B's 403B total parameters with 10 active per token across 60 layers, its roughly 807 GB of BF16 and F32 weights across about 188 shards, its Apache-2.0 licence against Super VL's OpenMDW-1.1, and the 1M-token context of Super VL against Intern-S2-397B's 256K text window with multimodal evaluation at 64K, with an OrcaRouter logo in the corner.
Engineering & Research

NVIDIA Nemotron 3.5 Super VL 120B vs Intern-S2-397B: two open vision models, one of which you can actually call today

Date Published

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

Both of these models are open-weight, both are multimodal, both are Apache-style or OpenMDW licensed for commercial use, and both are the kind of thing you download rather than subscribe to. The difference that decides most evaluations is not on either spec sheet. Intern-S2-397B, InternLM's 403-billion-parameter flagship released on Hugging Face on 13 September 2026, has a vendor-hosted API you can point an O​penAI-compatible client at the same day you read about it. N​VIDIA N​emotron 3.5 Super VL 120B A12B, uploaded to Hugging Face on 12 October 2026 and confirmed in N​VIDIA's announcement on the 15th, does not: N​VIDIA has published no endpoint and no per-token price for it on any path, including its own NIM microservice. So one of these you can evaluate this afternoon with a token, and the other requires an H200 or four H100s before the first prompt. That gap is worth more than any seven-point spread on a benchmark table.

The deployment floor, side by side

Start here, because it eliminates one of these models for most readers before quality is discussed at all.

• Total and active parameters — Intern-S2-397B 403B total with 10 active per token across 60 layers and 512 experts; N​emotron 3.5 Super VL 120B A12B 120B total with 12 active per token across interleaved Mamba-2, MoE and attention layers.

• On-disk weight size — Intern-S2-397B around 807 GB across roughly 188 BF16/F32 shards; Super VL's BF16 build is a fraction of that, with an NVFP4 build small enough to run on one DGX Spark.

• Minimum viable hardware — Intern-S2-397B needs a real multi-GPU fleet; Super VL needs 4× H100 or 2× GB200 in BF16, or one H200, one GB200, 2× H100 or one DGX Spark in NVFP4.

• Hosted access — Intern-S2-397B is on the vendor's own Intern API at chat.intern-ai.org.cn/api/v1 under the model id intern-s2-397b, documented as O​penAI-compatible and A​nthropic-compatible; Super VL has no published endpoint from anyone.

• Licence — Apache-2.0 for Intern-S2-397B; OpenMDW-1.1 for Super VL. Both permit commercial use.

• Context — 256K tokens for Intern-S2-397B, with multimodal evaluation at 64K; 1M tokens for Super VL.

• Video input — Super VL accepts MP4 up to two minutes, sampled as frames; Intern-S2-397B's documented modalities are text, image and raw signal data. Video is not among them.

Read that list twice and the shape becomes clear. Intern-S2-397B is the bigger, older, more scientific model with a hosted door. Super VL is the smaller, newer, more deployable model with a longer window and video, and no door at all yet.

What Intern-S2-397B was actually built for

InternLM describes the model as scaling along three axes, and the first is the technical claim worth understanding: vision-language pre-training performed directly on raw pages of scientific literature, jointly modelling symbolic semantics and visual relationships in a shared representation space, with no intermediate parsing step. A figure, a table, a spectrum and the sentence describing them are learned together rather than text being extracted and images attached afterwards.

That design shows up in the benchmarks InternLM chose to lead with, and it is why the two models' vision numbers are not interchangeable. Intern-S2-397B's card reports ChartQAPro 71.12 and MMMU Pro 81.68 and leads on science-specific scores — Biology-Instructions 55.71, Mol-Instructions 53.95, MolecularIQ 62.35, SciReasoner-1.5 63.28, TOMG-Bench 66.76. Super VL's card leads with OCR-Reasoning 65.08, OCRBench-v2 at 68.12 English and 59.15 Chinese, MMLongBench-Doc 62.82, CharXiv-RQ 80.82 and MathVision 78.16. Different tasks, different training distributions, and neither set is independently reproduced.

The third axis is the one that makes Intern-S2-397B unusual for an agentic workload: black-box agentic reinforcement learning in sandboxed environments, with the agent connected through several frameworks and explicitly aimed at long-horizon task ceilings. N​VIDIA's equivalent claim for Super VL is a post-training pipeline ending in RL with verifiable rewards and multi-teacher on-policy distillation, evaluated on SWE-bench and tau2-bench. Both are assertions of method rather than measurements of result.

The signal-input capability that neither competitor has

Intern-S2-397B accepts raw signal data — .npy, .csv and .wav — as a first-class input type alongside text and images, with a documented earthquake-event example detecting P-wave and S-wave arrival indices, and point-and-quantile forecasting through a forecast_horizon parameter. That is currently LMDeploy-only, and it is an inference-time feature rather than a benchmark result.

No model in this comparison does anything like it, and no closed frontier model does either. If your work involves sensor traces, seismic data or time-series forecasting that has to be reasoned about alongside images and text, Intern-S2-397B is not the better choice among these two — it is the only choice, and the 807 GB of weights are the entry fee.

Benchmarks, with the sourcing attached to each

Neither model has a clean independent evaluation, and the two vendor tables do not overlap on a single row, so what follows is a reading of claims rather than a ranking.

Intern-S2-397B, all vendor-reported by InternLM on its own harnesses (OpenCompass, VLMEvalKit, AgentCompass), unreproduced, with GPT-5.5 as the closed reference: MMLU Pro 89.77, HMMT-2026 93.56, AdvancedIF 71.57, MMMU Pro 81.68, ChartQAPro 71.12, TerminalBench 2.1 64.04, SWE-bench-Pro 68.54, SWE-bench-Multilingual 84.00, SciCode 47.63, SGLBench 51.89, FrontierScience-Olympiad 87.00 and FrontierScience-Research 55.00. Artificial Analysis had no Intern-S2-397B entry at release and the neutral leaderboards did not carry one either — that gap may have closed since, and it is worth checking before you rely on any of these.

N​emotron 3.5 Super VL 120B A12B, split between two sources. Vendor-reported by N​VIDIA on NeMo Gym: SWE-bench Verified 71.10, SWE-bench Multilingual 67.72, SWE-bench Pro 51.78, tau2-bench retail 69.24 and airline 76.50, BrowseComp 78.91, GraphWalks 89.71, GPQA-Diamond 89.52, HLE 31.79 without tools, OCRBench-v2 68.12 English, MMLongBench-Doc 62.82, Video-MME-v2 32.85. Panel-measured by Artificial Analysis, and this is the only independently produced number in this entire article: an Intelligence Index of 26.49, with 0.0 on Terminal-Bench 4.0 and a measured hallucination rate of 0.3585.

Two things deserve flagging in that second block. N​VIDIA's Terminal-Bench 2.1 figure appears as 66.70 on the model card and 60.67 in the technical report's summary table, with every other row in those documents agreeing — one of the two is wrong, and until the recipes are re-run neither is safe to quote. And the panel's 0.0 on Terminal-Bench 4.0 against a vendor-reported 64.04 on TerminalBench 2.1 for Intern-S2-397B, measured on a different suite by its own lab, is a reminder that terminal-agent scores are not comparable across harnesses even when the names look similar.

A two-column comparison scoreboard titled NVIDIA Nemotron 3.5 Super VL 120B vs Intern-S2-397B, with rows for Parameters, Context, Weight footprint, Licence, Hosted API, Video input and Independent score, carrying Super VL's 120B total and 12B active, 1M-token context, single-H200 NVFP4 serving, OpenMDW-1.1 licence, no published endpoint, MP4 video up to two minutes and its 26.49 panel index with 0.0 Terminal-Bench 4.0, against Intern-S2-397B's 403B total and 10 active across 512 experts, 256K text and 64K multimodal context, roughly 807 GB across about 188 shards, Apache-2.0 licence, vendor Intern API, no video input, and no panel entry at release.

Why the 1M-token window versus 256K is not a seven-figure win

The spec sheets say Super VL serves 1M tokens and Intern-S2-397B serves 256K, with multimodal evaluation at 64K. That looks like a rout for the N​VIDIA model until you attach the panel's long-context recall figure: 0.69 on AA-LCR. A model with a million-token window and a sub-one long-context score is a model that can ingest a million tokens and not necessarily do much with them at the far end.

Intern-S2-397B's 256K is documented at 256K for text reasoning and 64K for multimodal, and InternLM has published no long-context recall figure at all. So the honest position is that one model has a longer window and a poor measured recall on it, and the other has a shorter window and no measurement. If long-context recall is central to your workload, neither of these should be selected on the strength of its context length.

Availability, and how OrcaRouter fits when neither model is on our catalogue

Neither Intern-S2-397B nor N​VIDIA N​emotron 3.5 Super VL 120B A12B is served by OrcaRouter today, and this article will not imply otherwise. Intern-S2-397B is reachable through the vendor's own API and self-hosting with LMDeploy, vLLM or SGLang, with deployment recipes shipped in the model card. Super VL is self-hosted only — N​VIDIA's NIM microservice is the announced path, and no endpoint or price has been published for it.

What a gateway is worth when the models you want are not on it has a specific answer: it is where the rest of the pipeline already lives. Document triage that has to stay inside your network runs on your own Intern-S2-397B or Super VL deployment; the extraction, classification and summarisation calls around it run on a hosted model on the same key. The routing DSL is what keeps that boundary legible — expressed as a condition on the prompt rather than as a branch buried in application code — and automatic failover means the self-hosted half failing does not take the request with it. Zero markup matters here in a duller way: whatever the hosted half of your pipeline costs today, it costs that here, because provider list price is passed through unchanged.

The OrcaRouter models index page, showing the catalogue of more than two hundred models with provider names and per-million-token pricing.

Choosing between them

Choose Intern-S2-397B if you need to be running today without buying hardware, if your domain is scientific literature and figures, if raw signal or time-series input is part of the workload, or if a vendor-maintained hosted endpoint is worth more to you than control over the deployment. The 807 GB of weights are a real cost, but the hosted API means you can establish whether the model is good before you pay it.

Choose N​emotron 3.5 Super VL 120B A12B if video is an input, if one H200 or a DGX Spark rather than a fleet is what you have, if the model needs to sit inside your network, or if you intend to fine-tune — N​VIDIA published the post-training datasets and the NeMo Gym RL environments alongside the weights, and that is the more reusable half of the release. Run your own document set through the NVFP4 build before committing, and test long-context recall specifically rather than trusting the 1M-token figure.

Choose neither, for now, if what you actually need is a model with an independent evaluation you can lean on. Intern-S2-397B's numbers are all vendor-reported and its panel coverage is unclear; Super VL has exactly one third-party score and it is a single index figure. Two open vision models, roughly six weeks apart in release, both with unreproduced benchmark tables and neither on a neutral leaderboard in a way you can compare — the responsible move is to benchmark them on your own documents, and the vendor API is what makes half of that cheap.

What to watch: an independent panel entry for Intern-S2-397B, which would turn its entire card from claims into evidence; a price and an endpoint for Super VL, which would remove the H200 from the entry requirements; and a corrected Terminal-Bench figure from N​VIDIA, which would resolve the one place where a vendor's own two primary documents contradict each other.

The Hugging Face model card for NVIDIA Nemotron 3.5 Super VL 120B A12B, showing the model summary and the vendor benchmark table with the vision-language evaluation rows.

None of this is a reason to avoid the model, only a reason to know the history: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.