Nemotron 3.5 Super VL 120B-A12B review, pricing and limits

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, with context of up to 1M tokens and a median output speed of 436.2 tokens per second across hosts, as measured by Artificial Analysis. It suits teams that want one open model for document, screenshot and tool-calling agents instead of running a separate text model and vision model.

Nemotron 3.5 Super VL is NVIDIA's open vision-language model, scoring 89.52 on GPQA Diamond in NVIDIA's full-precision evaluation. A mixture-of-experts design with 120B total and 12B active parameters, it reads text, images and video, then reasons and calls tools from one checkpoint released under the OpenMDW-1.1 license.

Provider: NVIDIA · Family: Nemotron 3.5

More about NVIDIA on HokAI

Context window: 1,000,000 tokens · Max output: 262,144

Input modalities: text, image, video · Output: text, tool-calls

About Nemotron 3.5 Super VL 120B-A12B

Nemotron 3.5 Super VL is an open vision-language model from NVIDIA, released on October 15, 2026 alongside the edge model Nemotron 3.5 Bolt 74B-A9B. It is a mixture-of-experts (MoE) network that interleaves Mamba-2 state-space layers, MoE layers with 512 routed experts each, and attention layers, with a C-RADIO vision encoder and a connector that feed images and video frames into the language backbone. The backbone starts from Nemotron 3 Super Base, a lineage that includes the text-only 120B-A12B Nemotron 3 Super NVIDIA shipped on March 11, 2026, and was then distilled from Nemotron 3 Ultra. NVIDIA says training used about 26 trillion tokens, followed by roughly 3.8 million post-training samples and reinforcement learning across dozens of environments. It sits beside Nemotron 3.5 Lightning, the text-only high-volume model, as the multimodal tier of the NVIDIA model line for agents that must read pages as well as call tools.

On NVIDIA's BF16 model card the model scores 89.52 on GPQA Diamond, 71.10 on SWE-bench Verified, 51.78 on SWE-bench Pro, 60.67 on Terminal-Bench 2.1, 88.22 on LiveCodeBench v6, 70.86 on IFBench, 51.14 on OSWorld-Verified and 31.79 on text-only Humanity's Last Exam. The NVFP4 checkpoint stays close on most rows (89.90 on GPQA Diamond) with a larger drop on MMLU-ProX, 73.76 against 76.91. NVIDIA also reports 63.44 percent on Hermes Bench and 87.05 percent on LangChain agent evals, both ahead of the much larger Nemotron 3 Ultra. Every one of those figures is vendor-reported. The card sets the model against Qwen3.8-Flash-Next, GLM-5.3-Flash and DeepSeek-V4-Flash-0731, and on the headline coding and knowledge rows at least one of them scores higher. Artificial Analysis, an independent evaluator, gives the model 26 on its Intelligence Index v4.3.2 and a median time to first token of 0.82 seconds, and counts 170M output tokens to run the index.

NVIDIA specifies up to 1M tokens of context, shared by the prompt, images, conversation history, tool results and the reply, and notes that individual deployment recipes may use less to fit GPU memory. OpenRouter lists a 1,048,576-token window and up to 262,144 completion tokens. Long-context results, which matter against the other models with 400K or more context, are uneven: 89.71 on GraphWalks, above every peer in NVIDIA's table, but 68.62 on AA-LCR, below the 75.1 to 79.7 that those peers post.

Inputs are text, images and mp4 video up to two minutes (compare other video-input models), with documents supplied as page images plus text; output is text. Hosts such as OpenRouter expose function calling and JSON-schema structured output, and reasoning can be switched off, set to medium or left fully on through the chat template. On vision rows NVIDIA reports 80.82 on CharXiv reasoning, 85.60 on V*, 78.16 on MathVision and 62.82 on MMLongBench-Doc, and its technical blog calls the model one of the top open entries on the OCRBench v2 leaderboard. Video is the thin spot at 32.85 on Video-MME v2, and not every serving recipe accepts it: NVIDIA's NVFP4 recipes for DGX Spark and for SGLang list text and image only.

Weights ship as BF16 and NVFP4 checkpoints on Hugging Face. BF16 needs four H100s or two GB200s. The 4-bit NVFP4 checkpoint fits a DGX Spark, an H200, a pair of H100s or a GB200 under vLLM or SGLang, with TensorRT-LLM listed for the GB200. Community GGUF builds from Unsloth span 46.1GB at 1-bit to 132GB at 8-bit, and llama.cpp runs them. Hosted access comes from NVIDIA's free trial endpoint on build.nvidia.com, from providers on OpenRouter such as Together AI, and from Baseten; the pricing notes carry the rates. NVIDIA NIM packages the model as a container that NVIDIA says is refreshed every two weeks.

NVIDIA filtered training data with a Gemma-3 4B-based guard model trained on its Nemotron Content Safety Dataset v2, mitigates bias with supervised fine-tuning and RLHF, and measures it with the BBQ benchmark; its bias card says testing was limited to English and multilingual parity is not claimed. The model cards name no red-team partners and cite no responsible scaling policy. Foundation data was collected from 2013 to December 2025 across English, 19 other spoken languages and 43 programming languages, post-training focused on French, German, Italian, Japanese, Spanish and Chinese, and 10B tokens of Indic continued pre-training target Indian languages. The license, which puts it in the open-source group, grants use without restriction and places no limits on outputs, and NVIDIA published an EU public summary of the training content.

The model fits agents that read documents, charts and screenshots and then act. NVIDIA names claims review, receipt extraction, enterprise search and security triage as targets, and cites Scale AI, Ramp, AI21 Labs and Perplexity among organizations customizing or integrating it. It is a poor first pick for terminal-driven coding agents or browsing research, where Qwen3.8-Flash-Next and GLM-5.3-Flash post higher scores in NVIDIA's own table, and for plain-text pipelines, where Nemotron 3.5 Lightning is the smaller NVIDIA option. Version 1.0 is the only release so far, and the launch post's only roadmap item is a Bolt E4B edge model due October 23.

Pricing

NVIDIA sets no list price because the weights are open, so each host prices the model itself. Rates read on 20 October 2026: on OpenRouter the lowest was Crusoe at $0.10 input, $0.40 output and $0.08 cached input per 1M tokens, Together AI, DeepInfra and CoreWeave charged $0.15 and $0.50 with cached reads at $0.02, and Venice charged $0.25 and $1.25. [Baseten](/hub/services/baseten) lists $0.20 input, $0.04 cached and $0.50 output on its Model API. Artificial Analysis shows medians of $0.225 and $0.575 across providers. NVIDIA's free endpoint on build.nvidia.com is meant for prototyping. Self-hosting swaps token fees for GPU rental time; Unsloth's guidance for its smallest [llama.cpp](/hub/tools/llama-cpp) build is at least 64GB of RAM or unified memory.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Hybrid Mamba-2 MoE backbone: Interleaves Mamba-2, mixture-of-experts and attention layers, with 512 routed experts per MoE layer and a shared Multi-Token Prediction head that drafts tokens for built-in speculative decoding.
• Text, image and video in one model: A C-RADIO vision encoder reads images, document pages and mp4 clips up to two minutes, and the model answers in text.
• Three reasoning levels: The chat template switches reasoning off, to medium or fully on, so simple turns can skip the thinking tokens that agents otherwise pay for.
• NVFP4 build for small footprints: The NVFP4 checkpoint, trained with quantization-aware distillation, runs on one DGX Spark, one H200, two H100s or one GB200.
• Open weights, data and RL environments: Weights, post-training datasets and NeMo Gym environments are open, so teams can adapt it with supervised fine-tuning, LoRA or reinforcement learning.
• LiLiCorr drafter for DGX Spark: An optional speculative drafter that NVIDIA says generates tokens up to 2.3x faster than decoding without speculation on DGX Spark.

Pros

• One checkpoint covers documents, screenshots, short video and tool calls, so an agent does not pass context between a text model and a vision model.
• Several independent hosts already serve it, and the permissive license lets a team move to its own GPUs without renegotiating terms.
• NVIDIA reports it ahead of the larger Nemotron 3 Ultra on two agent evals, the closest evidence that distillation kept agent skill.
• Artificial Analysis rates its output speed well above the median for open-weight models of similar size.

Cons

• Behind Qwen3.8-Flash-Next on every vision row of NVIDIA's table except English OCRBench v2, and well behind it on terminal and browsing tasks.
• Weak on long banking-style tool dialogs: 18.88 on tau3-bench banking, against 44.5 for Qwen3.8-Flash-Next in the same table.
• Nearly every score is vendor-reported with NVIDIA's own test settings, and independent results so far come mainly from Artificial Analysis.
• The 1M-token window depends on the host, and long-document reasoning is the weaker side of its long-context profile.
• Verbose reasoning: Artificial Analysis counts 170M output tokens to run its index, which raises per-task cost on pay-per-token hosts.

Benchmarks

• IFBench: 70.9% vendor-reported · 20 Oct 2026 — How precisely the model follows detailed instructions, % passing.
• GPQA Diamond: 89.5% vendor-reported · 20 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Pro: 51.8% vendor-reported · 20 Oct 2026 — Harder, longer real-repository coding tasks, % solved.
• LiveCodeBench v6: 88.2% vendor-reported · 20 Oct 2026 — Recently published coding problems, % solved.
• OSWorld Verified: 51.1% vendor-reported · 20 Oct 2026 — Tasks completed by operating a real desktop, % solved.
• SWE-bench Verified: 71.1% vendor-reported · 20 Oct 2026 — Real GitHub issues fixed end to end, % solved.
• Terminal-Bench 2.1: 60.7% vendor-reported · 20 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
• Humanity's Last Exam: 31.8% vendor-reported · 20 Oct 2026 — Expert-written questions across many fields, % correct.
• AA Intelligence Index: 26 cited: Artificial Analysis · 20 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $0.14/M cited: Artificial Analysis · 20 Oct 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 436 tok/s cited: Artificial Analysis · 20 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Nemotron 3.5 Super VL 120B-A12B Official Page