NVIDIA Nemotron 3.5 Lightning Is Live on GMI Cloud: What Your Agentic System Was Missing

NVIDIA Nemotron 3.5 Lightning is now available on GMI Cloud with Day-0 support. From the moment NVIDIA flips the switch, you can call Nemotron 3.5 Lightning through GMI Cloud's OpenAI-compatible serverless API.

Two months ago, we brought Nemotron 3 Ultra to developers on Day 0, NVIDIA's 550B frontier-reasoning model for the hardest orchestration and planning calls. Nemotron 3.5 Lightning—a 30B MoE with just 3B active parameters, distilled from Ultra itself,—complements Ultra by handling the high-volume, always-on tasks that make up most of what an agent actually does.

Why Nemotron 3.5 Lightning Is Different

Here's the uncomfortable truth about agentic AI: most of the tokens your agents burn aren't spent on brilliant reasoning. They're spent on the repetitive middle: summarizing a PR, classifying an alert, enriching context, checking a policy rule, calling a tool, formatting a result. Routing every one of those steps to a frontier model is like dispatching a container ship to deliver a single package.

Nemotron 3.5 Lightning is purpose-built for that volume layer. Distilled from NVIDIA's frontier Nemotron 3 Ultra and developed with the Nemotron Coalition, it's trained specifically for popular agent harnesses, with leading out-of-the-box accuracy on coding, tool calling, instruction following, and multi-turn workflows. And it's fully open, so you can post-train it for your domain and own the result.

Up to 4x Higher Throughput

Speed is the headline. It delivers up to 4x higher throughput compared to other models in its class and completes tasks up to 30% faster. For always-on sub-agents, that speed compounds: more steps completed per time budget, more tasks finished per dollar.

Seeing It Build in Real Time

We ran Nemotron 3.5 Lightning through the GMI Cloud API and watched it build several small apps live, a bouncing DVD screensaver, a fireworks toy, a Matrix effect, a Pomodoro timer, and a 2048 style game. Each build finished in seconds and used only a few hundred tokens for reasoning, showing exactly what the throughput numbers above mean in practice for repetitive, high-volume agent tasks.

A Generational Leap in Accuracy

This Lightning  model is a major step up from the previous Nemotron 3 Nano and is one of the frontier models on Artificial Analsysi Intelligence Index benchmark.

That's more than double the coding score and nearly double the knowledge-work rating of its predecessor, alongside the strongest hallucination-resistance numbers in its comparison set. NVIDIA notes these are preliminary results from a model still in training, and final GA numbers are expected to improve further.

Architecture Highlights

Dflash is a speculative decoding method that uses a lightweight model to draft multiple future tokens in parallel. The Lightning model verifies the entire block of tokens at once and accepts the valid tokens, accelerating generation while preserving accuracy.

Multi-Token Prediction generates multiple future tokens per forward pass, cutting latency on the long structured outputs agents produce constantly: summaries, findings, tool arguments, reports.

1M-token context lets a small model do big-model things, holding full repositories, alert histories, or multi-turn session state without truncation.

One System of Models, One API

Nemotron 3.5 Lightning also fits into a broader multi-model agent strategy. With NVIDIA NeMo Switchyard, an open-source model routing library, agent tools can intelligently send each request to the right model for the task based on needs such as quality, latency, and cost, without requiring developers to rewrite their applications.

That's the architecture we've been betting on. GMI Cloud already serves Nemotron 3 Ultra for the frontier calls. Now Lightning covers the high-volume specialized calls, and because both live behind the same OpenAI-compatible endpoint, a router like Switchyard can move traffic between them without you changing a line of application code. That gives teams a way to optimize model selection across accuracy, latency, and cost throughout the whole workflow.

Why Run It on GMI Cloud

Day-0, Serverless, Zero Idle

Nemotron 3.5 Lightning is live on GMI Cloud's serverless inference API the day it launches. Pay per token with zero-idle billing, or reserve dedicated endpoints when your agent fleet needs guaranteed throughput.

OpenAI-Compatible, One-Line Integration

Point your existing SDK at GMI and go. If you're already running Ultra, Qwen, GLM, or Kimi on GMI Cloud, adding Lightning to your routing pool is a model-string change.

Built for Hybrid Deployments

Lightning is designed to run anywhere: RTX on a workstation, DGX Spark on a desk, or H100/Blackwell in the datacenter. Prototype against GMI Cloud's hosted endpoint, then post-train and deploy your customized checkpoint on dedicated GMI infrastructure when you're ready for production scale.

What You Can Build

Always-On Personal Agents

Long-running assistants that manage email, calendars, projects, and bookings need a model that's fast, cheap, and tireless. Lightning's token efficiency makes persistent background agents economically viable.

Software Development Sub-Agents

PR summarization, code classification, test triage: the targeted tasks that surround your frontier coding agent. Teams like CodeRabbit already use Nemotron models for context enrichment before handing off to larger models for review and verification.

Cybersecurity Operations

Alert enrichment, incident classification, log queries, indicator correlation, structured findings for analysts. High volume, tight latency budgets, exactly Lightning's lane.

Financial Services and Telecom Workflows

Document extraction, policy checks, risk-signal monitoring, network alarm triage, billing questions. Post-train Lightning on your domain and it becomes a specialist you own outright.

Get Started

curl https://api.gmi-serving.com/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer $GMI_API_KEY" \
  --data '{
    "model": "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello! What can you do?"}
    ],
    "max_tokens": 2048,
    "stream": true
  }'

Check out the full documentation at docs.gmicloud.ai

Frontier models get the headlines, but systems of models win in production. Nemotron 3.5 Lightning is the fast, open, customizable workhorse that makes the economics of always-on agents actually work, and GMI Cloud gives you Day-0 access to it on infrastructure built for exactly this.

Deploy Nemotron 3.5 Lightning on GMI Cloud | Read the Docs | Join the Community on Discord