LLMPrivate

NVIDIA Nemotron 3.5 Lightning 30B

NVIDIA's 30B open Mixture-of-Experts model optimized for fast, low-latency execution in agent workflows, with 1M context and speculative decoding.

Maker
NVIDIA
Modality
Text
Context
1,000K tokens
Open weights
OpenMDW License Agreement, version 1.1

Overview

What is NVIDIA Nemotron 3.5 Lightning 30B

NVIDIA Nemotron 3.5 Lightning 30B is an open-weight, Mixture-of-Experts (MoE) language model released by NVIDIA on August 11, 2026. With 30B total parameters and only 3B active per token, it's optimized for high-volume, low-latency execution in autonomous agent pipelines, supporting up to 1M tokens of context.

Running it privately on Venice

On Venice, you run NVIDIA Nemotron 3.5 Lightning 30B with full privacy — zero retention, no profiling, and no use of your prompts for training. The model runs with end-to-end function calling, web search, and reasoning capabilities, making it ideal for agentic workflows where sovereignty and uncensored execution matter. Since it's open weights, you can also inspect and customize the model within Venice's permissionless environment.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Highly efficient for agent execution: activates only 3B of its 30B parameters per token, reducing latency and cost.
  • Supports up to 1M tokens of context, ideal for long-running autonomous agents and complex agentic workflows.
  • Includes native speculative decoding (Multi-Token Prediction) for up to 4x faster output generation.
  • Open weights and permissive OpenMDW license allow for customization, auditing, and commercial use.
  • Optimized for tool calling, code generation, and structured reasoning tasks.
Limitations
  • Not designed for deep, single-shot reasoning: trails larger dense models like Qwen 3.6 35B-A3B in complex science and engineering tasks.
  • English and coding languages are strongest; support for Spanish, French, German, Italian, and Japanese is present but less robust.
  • Still requires significant GPU memory despite MoE efficiency — not suitable for low-end hardware.
  • No vision input: text-only model, limiting multimodal use cases.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
NVIDIA
Released
August 11, 2026
Architecture
MoE - Mamba-2 + MoE + Attention hybrid
Total parameters
30B (3B active)
Open/closed
Open weights
Context window
1,000K tokens
Max output
32.768K tokens
Capabilities
Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Private — zero retention
Available on Venice since
Aug 2026
License
OpenMDW License Agreement, version 1.1

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

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-lightning-30b-a3b",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.10 per 1M input tokens and $0.25 per 1M output tokens.

Input / 1M tokens
$0.10
Per 1M tokens
Output / 1M tokens
$0.25
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelMax contextOpen weightsPrice (in/out per 1M)Best for
NVIDIA Nemotron 3.5 Lightning 30B1M tokensYes$0.10 in · $0.25 out / 1MAgent execution
DeepSeek V4 Flash 07311M tokensYes$0.17 in · $0.35 out / 1MGeneral speed
Google Gemma 4 31B Instruct256K tokensYes$0.12 in · $0.36 out / 1MLocal fine-tuning
Claude Sonnet 4.61M tokensNo$3.60 in · $18 out / 1MBalanced reasoning

Fast, open MoE model optimized for agent workflows with 1M context.

Use cases

What it is good for

  1. 01Execution layer in agentic workflows: tool calls, result validation, sub-agent delegation.
  2. 02High-throughput coding agents that fix, test, and refactor code across repositories.
  3. 03Long-context data summarization and processing for compliance or research.
  4. 04Customizable AI workhorse for enterprises needing private, auditable agent models.
  5. 05Fast, low-latency inference in local or on-premise agent deployments.

Prompting

Getting better results

Use structured JSON output for reliable parsing in agent pipelines.

Leverage web search and tool calling for real-time data retrieval and actions.

Set temperature to 1.0 and top_p to 0.95 for optimal performance as recommended.

Break complex tasks into smaller steps — this model excels in iterative execution.

Use it behind a stronger planner model for best results in reasoning-heavy workflows.

Version history

Nemotron 3 Nano
2025

Predecessor small model in the family.

Nemotron 3.5 Lightning 30B
2026-08

Current — optimized for agent execution with MoE and 1M context.

FAQ

Frequently asked questions

It is a 30B-parameter open Mixture-of-Experts (MoE) language model by NVIDIA, optimized for fast, low-latency execution in autonomous agent workflows. It activates only 3B parameters per token and supports up to 1M tokens of context, making it efficient for high-volume tool calling and agentic tasks.

On Venice, it costs $0.10 per 1M input tokens and $0.25 per 1M output tokens. Pricing is transparent and usage-based, with no subscription required.

It is not open source, but it has open weights released under the permissive OpenMDW-1.1 license, allowing free use, modification, and commercial deployment without royalties.

Yes, it supports function calling and tool use natively, making it well-suited for agentic workflows that require interacting with external APIs, databases, or code execution environments.

It supports up to 1,000,000 tokens of context, enabling it to process extremely long documents, codebases, or conversation histories in a single pass.

Yes. On Venice, it runs with zero retention — your prompts are never stored, profiled, or used for training. This ensures full privacy and data sovereignty for enterprise and personal use.

NVIDIA Nemotron 3.5 Lightning 30B is better for high-volume, low-latency agent execution with open weights and lower cost. Claude Sonnet 4.6 offers stronger single-step reasoning and is more polished for general tasks but is closed, more expensive, and not customizable.

It supports English, Spanish, French, German, Italian, Japanese, and major coding languages. English and code are strongest; other languages are functional but less robust.

Run NVIDIA Nemotron 3.5 Lightning 30B privately

No prompt logging. No data used for training.