Research Vault

NVIDIA Nemotron 3.5 Super VL 120B-A12B

Properties1
tagshuggingface, image-text-to-text, multimodal, vision-language, moe, hybrid, mamba, coding, agentic, tool-calling, long-context, nvidia

4 min read

NVIDIA Nemotron 3.5 Super VL 120B-A12B

One-line: NVIDIA’s new flagship open model — a 120B/12B-active Mamba-2 hybrid MoE VLM with 1M context built for always-on agents and document work; released 2026-10-15 with weights + datasets + recipes. Artificial Analysis (independent) scores it the leading US open-weight model (Intelligence Index 26.5 vs Inkling-Small 25.7, Nemotron 3 Ultra 23), using ~60% fewer output tokens per task than its predecessor. ⚠ 248.9 GB BF16 (≈60 GB NVFP4) — 4×H100/2×GB200-class serving, not a single-consumer-GPU model.

⚠ Coding/agentic numbers below are lab-reported, collected via NVIDIA’s NeMo Gym harness (recipes published); treat cross-model gaps as indicative. Independent AA aggregates (Intelligence Index 26.5, NVFP4 ≈ lossless, ~630 tok/s @100k prompt on NVFP4) are called out separately.

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench88.22 (v6)lab — NeMo Gym; vs Qwen3.8-Flash-Next 94.1, DeepSeek-V4-Flash-0731 94.6, Inkling-Small 88.3, GLM-5.3-Flash 80.6
Aider polyglot—not published
SWE-bench Verified71.10lab — NeMo Gym; vs Qwen3.5-122B-A10B 71.9, DeepSeek-V4-Flash-0731 75.4, Qwen3.8-Flash-Next 78.7, GLM-5.3-Flash 79.10
SWE-bench Multilingual67.72lab; trails Qwen3.8-Flash-Next 77.10, GLM-5.3-Flash 78.30
SWE-bench Pro51.78lab; trails Qwen3.8-Flash-Next 60.37, GLM-5.3-Flash 59.70. Harness-level coding-agent results published as an image chart only
SciCode (subtask)47.13lab; vs DeepSeek-V4-Flash 49.3, GLM-5.3-Flash 50.05
DeepSWE [offline]24.65lab; trails DeepSeek-V4-Flash 42.3, Qwen3.8-Flash-Next 55.57, GLM-5.3-Flash 57.77
Terminal Bench 4 [offline]1lab; Qwen3.8-Flash-Next 20.71, GLM-5.3-Flash 15.15

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ²-bench (retail / airline / telecom)69.24 / 76.50 / 73.19lab; telecom trails badly (Qwen3.8-Flash-Next 90.4, GLM-5.3-Flash 98.5)
τ³-Bench Banking18.88lab; vs Qwen3.8-Flash-Next 44.5, GLM-5.3-Flash 34.8
ToolACE—not published
Terminal-Bench 2.166.70lab; vs Inkling-Small 53.4, DeepSeek-V4-Flash-0731 76.8, GLM-5.3-Flash 82.7, Qwen3.8-Flash-Next 85.8
GAIA—not published
BrowseComp78.00lab (offline 78); vs Qwen3.8-Flash-Next 89.0, GLM-5.3-Flash 86.6
WebVoyager [offline]77.72lab; vs Qwen3.8-Flash-Next 90.39, GLM-5.3-Flash 72.64
OSWorld-Verified51.14lab
MCP-Atlas—not published for Super VL
GDPval-AA v21211lab (Elo-style); vs GLM-5.3-Flash 1563, Qwen3.8-Flash-Next 1561
AutomationBench-AA [offline]32.44lab; vs Qwen3.8-Flash-Next 72.06, GLM-5.3-Flash 70.17
AA-Briefcase-Lite [offline]1137 | 1252 | 40lab (Elo pair + duration)

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro76.91 (MMLU-ProX)lab; trails Qwen3.5-122B-A10B 83.2, GLM-5.3-Flash 83.0
GPQA Diamond89.52lab; vs DeepSeek-V4-Flash-0731 89.6, GLM-5.3-Flash 91.2
MATH—not published
AIME 2025 / 2026—not published
HLE31.79 (text-only) / 29.20 (full)lab; vs Qwen3.8-Flash-Next 39.6, GLM-5.3-Flash 36.47
CritPt10.86lab; vs DeepSeek-V4-Flash 12.6
APEX shortlist88.36lab
AA-LCR v1.168.62lab; trails DeepSeek-V4-Flash 79.7, GLM-5.3-Flash 79.5
IFBench70.86lab
GraphWalks89.71lab; leads set (GLM-5.3-Flash 88.4, Qwen3.5-122B 86.1)
WMT24++84.80lab

Multimodal (VL specifics — why it exists)

BenchmarkScoreSource
CharXiv reasoning80.82 (84 w/ Python tool)lab; vs Qwen3.8-Flash-Next 85.8
OCRBench v2 (EN / ZH)68.12 / 59.15lab; EN edges GLM-5.3-Flash 61.8
MMLongBench-Doc62.82lab
MathVision78.16lab
V*85.60lab
OCR-Reasoning65.08lab
Video-MME v232.85lab; vs Qwen3.8-Flash-Next 45.1
BabyVision29.60lab

The card markets vision for “OCR, tables, charts, multi-page documents, and videos” — parse a chart, link it to a table elsewhere in the same document, summarize both without a separate image model.

Independent verification (Artificial Analysis, 2026-10-15)

  • Intelligence Index 26.5 — leading US open-weight model; edges Thinking Machines’ Inkling Small (25.7), beats Nemotron 3 Ultra (23) and Meta Muse Glimmer (17). Still trails UAE K2 Horizon 375B A23B (31) and China’s MiMo-V2.6-Pro (46) on open-weights intelligence.
  • Token efficiency: ~46k output tokens per Intelligence Index task — ~60% fewer than predecessor Nemotron 3 Super (114k).
  • NVFP4 ≈ lossless: NVFP4 scores within 0.1 points of BF16 on the Index (no component moves >2.6 pts).
  • Speed: ~630 tok/s output @100k-token prompts (pre-release DeepInfra NVFP4 endpoint) — fastest AA has measured among models scoring ≥20 on the Index at that prompt length; 1.8s time-to-first-token. EXO Labs (DGX Spark): no other tested model was both more accurate and faster.

Efficiency

  • Active params: 12B per token of 120B total (512 routed experts/MoE layer; hybrid Mamba-2 + attention)
  • Context length: 1,048,576 (1M) — prompt + vision + history + tool results share the window
  • VRAM: BF16 = 248.9 GB across 64 safetensors (verified). Deployment: 4×H100 (TP4) or 2×GB200 (TP2) for 1M ctx, vLLM (nemotron-3.5-super-vl; NeMo Gym recipes published) or SGLang (lmsysorg/sglang:v0.5.20). NVFP4 sibling (-NVFP4, ~60 GB-class) + -NVFP4-LiLiCorr drafter for Blackwell; -Base-BF16 for pretraining. Native MTP speculative head (3-5 draft tokens, no separate drafter). ⚠ Not local-single-GPU; wait for community GGUF/Unsloth quants (already appearing) for anything smaller.
  • Runs on: DGX Spark / DGX Station / RTX PRO / RTX, Jetson (edge), data centers; NIM microservice w/ biweekly updates. Optimized for NVIDIA hardware throughout.

What makes it notable

NVIDIA’s successor jump: Nemotron 3 Super (a text model) → a 3.5-generation multimodal agent model that AA scores 2× its predecessor’s intelligence, #1 among US open weights. It was distilled from Nemotron 3 Ultra and NVIDIA claims agent-harness parity-or-better vs Ultra at ~1/5 the parameters (their own Hermes + LangChain agent tests). The interesting trade-off for this user: it is not a local model (248.9 GB BF16 / ~60 GB NVFP4), but its token efficiency (~60% fewer tokens per task) and throughput (~630 tok/s NVFP4) target exactly the long-running-agent workload. On the card’s cross-comparisons it beats Qwen3.5-122B and Gemma-4-26B-A4B on agentic rows, trails Qwen3.8-Flash-Next / DeepSeek-V4-Flash / GLM-5.3-Flash on most coding rows. Released alongside datasets (post-training v3, pre-training, Indic 2.83T tokens) and recipes. Sister model: nvidia—NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 (74B/9B, 24GB-local). Watch for: .eval_results-style machine-readable evals, community quants, and whether SWE-bench Pro harness spread matches the lab chart.

See also

Graph View

Backlinks