Wait Which Model?
← Back to directory
NVIDIA

Nemotron 3.5 Super VL

Released Oct 15, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
United States
Type
LLM
Modality
Multimodal
Context window
1M
Max output
—
Speed
422 tok/s · 0.80s to first answer token
Price ($/MTok in / out)
$0.10 / $0.40
Cost per task
$0.19
Open weights
Yes
News1 story
Strengths & weaknesses

Strengths

  • Text, images, documents and short video go through one checkpoint, so screenshot- and chart-heavy agent steps need no separate vision model
  • Streams fast enough that multi-step agent loops feel interactive, and context-reuse serving keeps attention and Mamba state across repeated agent calls
  • Ships base, post-trained and NVFP4 checkpoints with datasets and recipes, so it can be genuinely retrained — partners have already post-trained it for finance, security and receipt extraction

Weaknesses

  • Verbose on reasoning work — it generates well above the median token count to reach an answer, which eats into the per-token price advantage
  • The full window needs datacentre hardware: NVIDIA's DGX Spark recipes cap context far lower and serve one sequence at a time
  • Video input stops at two minutes and NVIDIA's multimodal examples take one image and one video per request

120B total / 12B active hybrid Mamba-2–Transformer MoE with a C-RADIO vision encoder, distilled from Nemotron 3 Ultra and released 2026-10-15 alongside Nemotron 3.5 Bolt. Benchmarks are NVIDIA's own BF16 model-card figures (GPQA Diamond 89.52, SWE-bench Verified 71.10, SWE-bench Pro 51.78, Terminal-Bench 2.1 60.67, HLE text-only 31.79 — 29.20 on the full set — and GDPval-AA v2 1211, lab-reported); the NVFP4 checkpoint scores within a point on most rows but drops on DeepSWE and WebVoyager. The card reports MMLU-ProX 76.91, not MMLU-Pro. NVIDIA publishes no first-party per-token price — $0.10/$0.40 is OpenRouter's listing (a logged free NVIDIA endpoint also exists), while Artificial Analysis' provider median is $0.225/$0.575. Artificial Analysis Intelligence Index 26 (v4.3.2) at $0.19 per task, 421.7 tok/s and 0.80s to first token (provider medians, reasoning variant, no effort level named); it generated about 170M tokens running the index against a 100M median, the basis for the verbosity note. No max output limit is published (examples use 16,384 with reasoning sharing the budget); the card gives a 2013–December 2025 data-collection period rather than a knowledge cutoff, so knowledgeCutoff is null. Distillation from Ultra is lineage, not a replacement claim, so predecessorId is null.

Benchmarks6 of 10 reported

GPQA Diamond

89.52%

SWE-bench Verified

71.1%

SWE-bench Pro

51.78%

Terminal-Bench 2.1

60.67%

Humanity's Last Exam

31.79%

GDPval-AA v2

1211

MMLU-Proretired

—

AIME (latest)retired

—

LMArena Elo

—

ARC-AGI-2

—
For developers

API model strings

  • nvidianvidia/nemotron-3.5-super-vl-120b-a12b
  • openrouternvidia/nemotron-3.5-super-vl

Licence

OpenMDW-1.1 · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor