Model
Model explorer

Nemotron 3.5 Super VL

OPEN
NVIDIA · Nemotron 3.5 family · released Oct 15, 2026

NVIDIA's open vision-language Nemotron, released October 15, 2026 alongside Nemotron 3.5 Bolt: a 120B-total / 12B-active hybrid Mamba-Transformer MoE with a C-RADIO vision encoder and a 1M-token context, under OpenMDW-1.1, in BF16 and NVFP4. Served through NVIDIA NIM and OpenRouter; NVIDIA sets no first-party token price. NVIDIA's NeMo Gym table puts it below Qwen3.8-Flash-Next and GLM-5.3-Flash on most agentic rows (Terminal-Bench 2.1 66.7 vs 85.8 and 82.7). NOT FILED: rows marked [offline] in the table (BrowseComp offline, WebVoyager, AA-Briefcase-Lite, AutomationBench-AA, GDP.pdf, Terminal Bench 4, DeepSWE) are non-standard offline runs; GDPval-AA v2 (1211) is on AA's re-based scale; HLE text-only (31.79) is a subset (the full-set 29.20 is filed); MMLU-ProX, AA-Omniscience, CritPt, APEX shortlist, AA-LCR, GraphWalks, LMArena v3 proxy, WMT24++, BabyVision, MMLongBench-Doc, V*, OCR-Reasoning, OCRBench v2 and Video-MME v2 have no matching corpus column.

ReasoningCodingVisionFunction callingTool useAgentic
2076.9
Elo · rank #87
Parameters
120B
Active params
12B (MoE)
Context
1M tokens
Architecture
Hybrid Mamba-2 / Transformer MoE (512 routed experts per MoE layer), 120B total / 12B active, with a C-RADIO vision encoder and multi-token prediction; 1M-token context
License
OpenMDW License 1.1
Languages
—
API price (in/out)
No hosted API
Modalities
text · vision · video
Benchmark results
Bar shows position within the tracked field; marker = field best
BrowseCompAgents78.0%#22
best: Kimi K3.1 Pro · 92.1%
CharXivVision80.8%#17
best: Qwen3.8-Flash-Next · 90.6%
GPQA DiamondReasoning89.5%#39
best: GPT-6 Astra · 96.0%
Humanity's Last ExamReasoning29.2%#63
best: Kimi K3.1 Pro · 70.0%
IFBenchReasoning70.9%#33
best: MiniMax M3 · 83.0%
LiveCodeBenchCoding88.2%#16
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MathVisionVision78.2%#13
best: Seed 2.1 Pro · 92.6%
OSWorld-VerifiedAgents51.1%#30
best: Claude Fable 5 · 85.0%
SciCodeCoding47.1%#4
best: Step 5 Preview · 58.9%
best: Claude Opus 5.5 · 93.9%
SWE-bench ProCoding51.8%#51
best: Claude Opus 5.5 · 89.9%
SWE-bench VerifiedCoding71.1%#73
best: Claude Opus 5 · 97.0%
τ²-Bench AirlineAgents76.5%#2
best: Trinity-Large-Thinking · 88.0%
τ²-Bench RetailAgents69.2%#11
best: Claude Sonnet 4.6 · 91.7%
τ²-Bench TelecomAgents73.2%#28
best: Claude Opus 4.6 · 99.3%
τ³-BankingAgents18.9%#7
best: Step 5 Preview · 42.5%
Terminal-Bench 2.0Coding66.7%#40
best: DeepSeek-V4.1-Flash · 90.6%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
NVFP4
Fine-tune it
Permissive
QLoRA81.6 GB1× H200 141GB
LoRA255.6 GB2× H200 141GB
Full fine-tune1926.0 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $62.06 (1× H200 141GB)
API price weights · each benchmark row carries its own source badge (see methodology)