Every step an AI agent takes — gathering context, reasoning, calling tools, acting — is a language model call. For always-on agents running high-volume workloads, model speed and cost per step are important factors alongside accuracy, reliability, security, and specialized task-specific performance.

NVIDIA Nemotron 3.5 Lightning is built for exactly that regime, and it is available on DeepInfra's serverless API from day zero — no waitlist, no GPU provisioning, no setup. Point your existing OpenAI-compatible client at nvidia/NVIDIA-Nemotron-3.5-Lightning and you can start sending requests immediately.

What is Nemotron 3.5 Lightning?

Nemotron 3.5 Lightning is NVIDIA's new open model for powering always-on agents to complete specialized tasks fast. It's the fastest open model in its class, delivering up to 4x higher throughput and up to 30% faster task completion for high-volume agentic workflows.

It's a 30B-parameter hybrid Mixture-of-Experts model with 3B active parameters per token, a context window that scales to 1M tokens, DFlash speculative decoding, and multi-token prediction — a text-in, text-out model distilled from NVIDIA frontier Nemotron 3 Ultra and trained on open datasets. Developed with the Nemotron Coalition, it's built for popular agent harnesses and designed for customization: organizations can post-train it for their own tools, workflows, and policies to achieve leading accuracy on specialized tasks, while keeping full control of the weights.

Architecture built for agent economics

The hybrid MoE design activates only 3B of the model's 30B parameters per forward pass, so compute per token is close to a 3B dense model while capacity remains that of the full 30B. DFlash speculative decoding and multi-token prediction accelerate generation of the long, structured outputs agents produce — tool calls, JSON, multi-step plans — contributing to the up to 30% faster task completion. And the 1M-token context supports long-running, multi-turn workflows without aggressive truncation of history, retrieved documents, or tool outputs.

The numbers

On Artificial Analysis's output-speed measurements, Nemotron 3.5 Lightning lands at roughly 4x the output speed of comparable open models in the 25–35B class, placing it alone in the high-speed, high-intelligence quadrant among its peers.

On accuracy, the model is trained for agentic tasks and popular agent harnesses, with leading results on the two capabilities that matter most for multi-step agent workflows:

• Agent productivity (PinchBench): 86.5% — ahead of same-class models like Gemma 4 26B (72.3%) and its predecessor Nemotron 3 Nano (62.8%).
• Knowledge reliability (AA-Omniscience Non-Hallucination): 69.9% — leading its class (next best: 50.3%), a key metric for agents that act autonomously on their conclusions.

Where it fits

Agentic systems increasingly route each workflow step to the right model: frontier models for complex reasoning and orchestration, fast specialized models for high-volume domain-specific steps. NVIDIA NeMo Switchyard is a configurable routing library that gives teams multiple algorithms for automating model selection according to their own criteria. Depending on the models, routing policies, and performance requirements a team configures, Nemotron 3.5 Lightning serves as the high-throughput, specialized workhorse in that system for workloads such as personal agents (email, calendar, projects, bookings), financial services (document extraction, policy checks, risk monitoring), cybersecurity operations (alert enrichment, incident classification, log queries), telecom (alarm triage, network optimization, billing), and retail (catalog enrichment, fulfillment exceptions, order and returns support).

Quickstart

DeepInfra's API is OpenAI-compatible — one base-URL change and you're running Nemotron 3.5 Lightning:

from openai import OpenAI

client = OpenAI(
    api_key="$DEEPINFRA_TOKEN",
    base_url="https://api.deepinfra.com/v1/openai",
)

response = client.chat.completions.create(
    model="nvidia/NVIDIA-Nemotron-3.5-Lightning",
    messages=[
        {"role": "system", "content": "You are a security operations assistant."},
        {"role": "user", "content": "Triage this alert: multiple failed SSH logins "
                                    "from 203.0.113.42 across 14 hosts in 3 minutes, "
                                    "followed by one successful login on db-prod-2."},
    ],
    max_tokens=1024,
)

print(response.choices[0].message.content)
copy

Streaming, tool calling, and JSON mode work as with any OpenAI-compatible endpoint.

Pricing

Nemotron 3.5 Lightning on DeepInfra is priced at $0.05 per 1M input tokens and $0.20 per 1M output tokens — serverless, pay per token, with no minimums, no idle cost, and no GPU instances to manage. Live pricing is on the model page.

Your data stays yours

Nemotron 3.5 Lightning's positioning centers on model control, and DeepInfra's inference layer matches it:

• We don't store your prompts or responses. Requests are processed and returned, not retained.
• We never train on your data. Inputs and outputs are not used to train or fine-tune any model.
• We're SOC 2 and ISO 27001 compliant — the controls behind those statements are independently audited.

Get started

We've been working with early-access checkpoints ahead of the release, so inference is optimized and endpoints are live from the moment the model is public. Starting today you can benchmark Nemotron 3.5 Lightning with your real workloads on production-ready infrastructure:

• Try it in the browser on the model page
• Read the docs for the OpenAI-compatible API at deepinfra.com/docs
• Read NVIDIA's announcement on the NVIDIA blog