NVIDIA

Nemotron 3.5 Super VL

NVIDIA's open 120B/12B multimodal agent model under OpenMDW-1.1, launched 2026-10-15. One independent tracked score so far: Artificial Analysis's Humanity's Last Exam at 30.72 (checked 2026-10-17). Terminal-Bench 4.0 is not recorded — AA's field reads 0, which this site does not treat as a measured score.

Nemotron 3.5 Super VL benchmarks and pricing, every number sourced: 1 tracked Nemotron 3.5 Super VL benchmark score (1 independently run, 0 still resting on a vendor’s own claim), priced at $0.23 per million input tokens and $0.57 per million output.

Nemotron 3.5 Super VL architecture: Hybrid Mamba-Transformer LatentMoE, multimodal (text + image + video); 120B total parameters (12B activated per token); 1M-token context window.

Released: 2026-10-15
License: open-weights
Context window: 1M tokens
Knowledge cutoff: Not disclosed
Verified: source checked 2026-10-17
Parameters: 120B (12B active)
Architecture: Hybrid Mamba-Transformer LatentMoE, multimodal (text + image + video)

Nemotron 3.5 Super VL’s verified record

Nemotron 3.5 Super VL’s most-compared rival is Beam 1.0: 1 trail across their 1 shared comparison. Nemotron 3.5 Super VL is priced at $0.23/$0.57 per 1M tokens (in/out) vs Beam 1.0’s $0.70/$2.10.

Against the 39 head-to-head comparisons Nemotron 3.5 Super VL shares with other tracked models:

38 real gaps0 inside the band1 not called

A gap counts for Nemotron 3.5 Super VL only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Nemotron 3.5 Super VL trails on 38 of them.

HLE · no tools Reasoning

±2 is noise

Behind

Beam 1.0 −3.9 · Mistral Large 4 −4.3 · Qwen3.8-Flash-Next −7.3 · GPT-6 Luna −7.8 · MiniMax M3 −8.3 · DeepSeek V4.1 Flash −8.5 · GLM-5.3-Flash −9.2 · DeepSeek V4 Pro (0813) −10.3 · GLM-5.3 −11.6 · Qwen3.8-Max −12.4 · Grok 4.7 −12.4 · Muse Spark Open −13.2 · Claude Haiku 5.5 −13.7 · Step 5 Preview −15.8 · Kimi K3 −16.2 · Gemini 3.1 Pro Preview −16.3 · Gemini 3.8 Flash −17.1 · Muse Spark 1.3 −18.0 · MiMo-V2.6-Pro −18.6 · Muse Wonder −19.4 · GPT-6.1 Sol −22.2 · GPT-6 Astra −24.0 · Claude Sonnet 5.5 −24.3 · Gemini 4 Argon −26.4 · Claude Fable 5.1 −28.4 · Claude Opus 5.5 −30.7 · 12 previous: DeepSeek V4 Flash (0731) −7.9 · GPT-5.6 Luna −8.8 · GLM-5.2 −10.4 · Claude Sonnet 5 −10.6 · Grok 4.6 −12.2 · Muse Spark 1.2 −14.7 · Gemini 3.7 Flash −17.2 · GPT-6 Sol −17.2 · Claude Opus 4.8 −18.0 · GPT-5.6 Sol −18.8 · Claude Opus 5 −24.2 · Claude Fable 5 −24.8

Unverified

1 model — vendor-reported on one side

Nemotron 3.5 Super VL API pricing

$0.23 in / $0.57 out per 1M tokens — official pricing source

What Nemotron 3.5 Super VL costs per job

Nemotron 3.5 Super VL cost for three reference workloads, computed from its list rates
Workload	Tokens in / out	Cost
One long chat turn	100K / 10K	$0.028
A codebase review	1,000K / 100K	$0.282
A day of agent work	10,000K / 1,000K	$2.83

Computed from Nemotron 3.5 Super VL’s list rates above — cache discounts and batch tiers are not applied.

Nemotron 3.5 Super VL benchmark scores

Nemotron 3.5 Super VL benchmark scores, provenance, and source links
Benchmark	Score
HLE(no tools)^[1] Reasoning · ±2 is noise	30.7artificialanalysis.ai · 2026-10-17

Who ran these numbers: 1 of 1 independent — artificialanalysis.ai (1).

HLE: AA's own run, board row 'Nemotron 3.5 Super VL' (reasoning; no effort-tier suffix): hle 0.307228915662651 = 30.72, no tools. Read from the Nemotron 3.5 Super VL page on 2026-10-17. The same object's terminalBench40 field is 0 — not imported as a Terminal-Bench 4.0 score.

Notes on the record

Released 2026-10-15 (NVIDIA blog "Open From Data Center to Desktop", checked 2026-10-17). Open weights on Hugging Face as NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (BF16) with an NVFP4 build, under OpenMDW-1.1. Specs: 120B total / 12B active, multimodal vision-language, up to 1M context. Bolt-family edge models shipped alongside it; they are not separate records here.

Hosted price on Artificial Analysis is $0.225 / $0.575 per 1M tokens (checked 2026-10-17). Serving the Nemotron weights yourself has no token fee.

Independent score from AA's nemotron-3-5-super-vl page (embedded data, checked 2026-10-17): Humanity's Last Exam 30.72. AA-AnalystAgent is null. The same object's terminalBench40 field is the number 0 — not null. This site does not create a Terminal-Bench 4.0 row from a literal zero without a published run explanation; the gap is noted here instead of silently imported. Intelligence Index on AA reads about 26.

Compare with

Beam 1.0 Muse Spark Open Muse Wonder Claude Haiku 5.5 any model…

FAQ

What is Nemotron 3.5 Super VL?

NVIDIA's 2026-10-15 open multimodal model: 120B total / 12B active, OpenMDW-1.1 weights, aimed at always-on agents that read documents and use tools. It is separate from Nemotron 3 Ultra and from the smaller Bolt edge models announced the same day.

Is Nemotron 3.5 Super VL open source?

Open-weights under OpenMDW-1.1 (Artificial Analysis and NVIDIA's model card, checked 2026-10-17). Weights are on Hugging Face; NVIDIA also published training recipes and datasets. OpenMDW is not identical to Apache 2.0 — read the license before commercial deployment.

How much does Nemotron 3.5 Super VL cost?

Running Nemotron 3.5 Super VL from the Hugging Face weights has no token price. Artificial Analysis lists a hosted rate of about $0.23 input / $0.58 output per million tokens (checked 2026-10-17). OpenRouter and NVIDIA NIM are among the hosts NVIDIA named at launch.

Has Nemotron 3.5 Super VL been independently benchmarked?

Thinly, on boards this site tracks. Artificial Analysis reports Humanity's Last Exam 30.72 (checked 2026-10-17). AA-AnalystAgent has no row. Terminal-Bench 4.0 is not recorded here because AA's embedded field reads 0, which is not treated as a verified run.

What is Nemotron 3.5 Super VL good at?

NVIDIA positions it for multimodal agent workflows — documents, charts, tool use — and for post-training on private data. The only independent tracked score here is HLE 30.72, so capability claims beyond that stay vendor-sourced until more boards fill in.

How does it compare with Muse Spark Open?

Both are US open-weight releases from mid-October 2026. On Artificial Analysis's Intelligence Index, Muse Spark Open is about 39 and Nemotron 3.5 Super VL about 26; on HLE, Open's xhigh 43.88 beats Nemotron's 30.72. Those are AA figures, not a composite ranking from this site.