NVIDIA's open 120B-parameter MoE vision-language model for agentic workflows — reasoning, tool use, document and video understanding at 12B active parameters.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia-nemotron-3-5-super-vl-120b-a12b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Overview
NVIDIA Nemotron 3.5 Super VL 120B is an open-weight vision-language model released in October 2026. Its mixture-of-experts architecture activates 12B of 120B parameters per token, combining reasoning, coding, tool use, and text, image, document, and video understanding in a single model built for specialized agents.
On Venice, Nemotron 3.5 Super VL runs under the private tier with zero retention — your prompts, images, and tool calls are never stored, profiled, or used for training. The weights are open under OpenMDW-1.1, so you get a permissionless, self-hostable model without handing your workflow data to an ad-funded platform. Note that Venice serves this model with standard infrastructure rather than TEE isolation or end-to-end encryption.
Agent quickstart
The API is OpenAI-compatible: change the base URL and the model id and existing client code works unchanged.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia-nemotron-3-5-super-vl-120b-a12b",
"stream": true,
"messages": [{ "role": "user", "content": "Draft the release note." }]
}'curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia-nemotron-3-5-super-vl-120b-a12b",
"messages": [{ "role": "user", "content": "Find the rate limits." }],
"tools": [{
"type": "function",
"function": {
"name": "search_docs",
"description": "Search the API documentation.",
"parameters": {
"type": "object",
"properties": { "query": { "type": "string" } },
"required": ["query"]
}
}
}],
"tool_choice": "auto"
}'import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["VENICE_API_KEY"],
base_url="https://api.venice.ai/api/v1",
)
resp = client.chat.completions.create(
model="nvidia-nemotron-3-5-super-vl-120b-a12b",
messages=[{"role": "user", "content": "Build without permission."}],
)
print(resp.choices[0].message.content)Specifications
Assessment
Use cases
Prompting
Toggle reasoning with the chat template setting: enable_thinking=true for hard multi-step tasks, false for fast extraction and classification.
Match sampling to modality — temperature 1.0 / top_p 0.95 for text tasks, temperature 0.6 / top_p 0.95 for vision inputs.
Pass multiple images in one request when comparing documents or frames instead of looping one image per call.
Define a JSON schema for structured output when the agent's result feeds another tool — it's supported natively.
Register tools with parallel_tool_calls enabled so multi-step workflows resolve in fewer turns.
Reuse a long system prompt across calls to hit the $0.13 cached-input rate instead of re-paying full input price.
Alternatives
| Model | Best for | Context | Open weights | Price (Venice) |
|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Super VL 120B | Multimodal agentic workflows | 262K tokens | Yes | $0.25 in · $1.25 out / 1M |
| Gemini 3.8 Flash | Fast everyday multimodal tasks | 1M tokens | No | $0.94 in · $4.69 out / 1M |
| Kimi K3 | Long-context reasoning and coding | 1M tokens | Yes | $3.75 in · $18.75 out / 1M |
| Qwen 3.7 Plus | Balanced general-purpose chat | 1M tokens | No | $0.50 in · $2 out / 1M |
| Google Gemma 4 31B Instruct | Budget open-weight tasks | 256K tokens | Yes | $0.12 in · $0.36 out / 1M |
Pick Nemotron 3.5 Super VL 120B when an agent has to read documents, images, or video, call tools, and finish multi-step workflows cheaply on open weights. For pure long-context reasoning at any price, Kimi K3's 1M window is the alternative.
Pricing
Billed per token on Venice: $0.25 per 1M input tokens and $1.25 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Getting a key
Create a key in API settings. Nothing else is required to start.
Export it as VENICE_API_KEY so the snippets above run unedited.
Point an existing OpenAI client at https://api.venice.ai/api/v1. The scheme is part of the value: an OpenAI client given a bare host does not resolve it.
Pass nvidia-nemotron-3-5-super-vl-120b-a12b as the model and send the request.
FAQ
It's NVIDIA's open vision-language model for specialized agents, released October 15, 2026. It pairs a Mamba-2–Transformer hybrid mixture-of-experts (120B total, 12B active parameters) with a C-RADIO vision encoder, handling reasoning, coding, tool use, and text, image, document, and video understanding in one model.
Venice bills it per token: $0.25 per 1M input tokens, $1.25 per 1M output tokens, and $0.13 per 1M for cached input. There is no subscription requirement — you pay for what you use.
The weights are open under the OpenMDW-1.1 license, which permits commercial and non-commercial use, and NVIDIA also publishes training data and recipes. Using the hosted model on Venice is paid per token, though new accounts get free daily credits to try it.
Yes — function calling is a core design target, including parallel tool calls in a single turn. It also supports structured output via JSON schema, web search grounding, and a toggleable reasoning mode, making it suited to multi-step agentic workflows.
Yes. It accepts text, images (multiple per request), and video as input, and outputs text. Document understanding — tables, charts, and page layouts — is one of its headline use cases, so it can replace a separate vision-language model in your stack.
Nemotron is the better fit for agentic workflows that need tool calling, document and video understanding, and open weights, at roughly a quarter of Gemini 3.8 Flash's input price. Gemini offers a 1M context window and closed-model polish, but you can't self-host or audit it.
Venice runs it under its private tier with zero retention — prompts, uploads, and outputs are not stored, profiled, or used for training. Note that this model is not served in a trusted execution environment and traffic is not end-to-end encrypted, so privacy rests on Venice's no-logging policy.
Venice serves it with a 256K-token context window and up to 32K output tokens per response. The architecture supports up to 1M tokens in NVIDIA's own deployments, so the Venice deployment is a practical subset of the model's maximum.
No. It ships with NVIDIA's standard safety alignment, so it will decline some prompts that Venice's uncensored models answer. If unfiltered responses matter more than multimodal agentic capability, choose one of Venice's uncensored models instead.
One key, free to start, no credit card.