Nemotron 3.5 Lightning Results

Your early-access path

Claim your name on local.ai

Refer someone else to unlock full access. Five qualified referrals earn VIP status.

Start your path

1

Step 1 of 3

Claim your name

Start here

ClaimFull accessVIP

Highlights

Most Intelligent Model

local.ai intelligence index

that runs on a

Fastest Model

Tasks per hour

that runs on a

Best Value Hardware

Local.ai Value Index

What should I run?

Intelligence / Speed Frontier

Pick hardware to compare measured intelligence and task time.

Select hardware to redraw the model Pareto frontier.

$$$4.7K

128GB

145/h

$$$5.2K

128GB

129/h

$$$$6.7K

128GB

161/h

$$$3.7K

64GB

107/h

$$$$6.8K

96GB

137/h

$$$2.7K

48GB

79/h

$$$3.5K

24GB

350/h

$$$3.7K

32GB

545/h

$$$$9.0K

48GB

347/h

$$$$$13.0K

96GB

527/h

Recommended for DGX Spark· NVFP4

5.36× faster than the most intelligent option while retaining 78% of its Intelligence Index score.

Real agent tasks

You don't always need a frontier model

Watch the same task run on both sides, replayed at 10x speed.

TaskAnimated car sceneLocal modelQwen 3.6 35B A3B

The task

A car driving through a surreal landscape, in one self-contained HTML file.

Both finish the job — Qwen 3.6 35B A3B is 21s faster and costs $0, while Claude Opus 4.8 charges $0.41 per run.

Qwen 3.6 35B A3B

Local

Running on DGX Spark

gary@local-ai website % h

10x replay · 3 steps · 3 calls · paused

WaitingPrefill · Step 1/3 · Write artifact.html10× speed

prefilldecodetool call29k in · 8.2k out

0s / 1m53s · $0working

Claude Opus 4.8

API

Running via API

gary@local-ai website % h

10x replay · 6 steps · 6 calls · paused

WaitingPrefill · Step 1/6 · Write artifact.html10× speed

prefillreasoningdecodetool call68k in · 10k out

0s / 2m14s · $0.41 per runworking