Palettemodels
»Connect
Palette · Seoul and Bengaluru · 2026

Korea's #1 on-device AI. Built by Palette.

…

»Explore the models»Connect
Measured againstKakao KananaLG EXAONENaver HyperCLOVA XAlibaba QwenSeedreamGPT-ImageCohere
82.2 %
…
505 MB
…
16
…

Open weights, Apache 2.0

…

Korean enterprise work, not trivia

…

Measured in public

…

Device ladder

How far down the device ladder do Palette models reach?

From a browser tab to a server rack: which Palette model runs at each tier today, with the file you actually download.

Phone, offline
50–800 MB · no server, no signal needed · Android and desktop browsers
Palette Island 12L · 54 MBPalette Island-X 0.8B · 505 MBPalette Island Pocket: mobile offline LLM
Laptop
8–64 GB · Apple silicon, one GPU
Palette Island-X 0.8B · toolsPalette STT v1c · 3.4 GBPalette Desktop
Workstation
24–80 GB · one H100 class card
Palette-K-Midm 11.5B · 23 GBPalette-K-14B · 29 GBPalette-K-Image 8B-MoT · 36 GB
Server
80 GB+ · multi-GPU, hosted API
Palette-K-Doc pipeline · 27BPalette-K-Video · H3 33B, APIModal, KR/SG regions
Footprint

Parameter count is not the same thing as RAM.

»See the models

…

Benchmarks

Every model, the number it was measured to, next to what it was measured against.

Hover a bar for the exact figure. White bars are Palette models. Each chart names its benchmark, scorer, item count and date.

The family

A family of models. One engineering discipline.

»All models and specs
Specs

All models and specs

ModelBaseParamsFileBenchmarkScoreLicenceStatus

…

Research

Papers, benchmarks and the sweeps behind every claim.

Published work links out. Reports in preparation already have their numbers on this page.

The stack

One pipeline from Korean data to the device in your pocket.

Palette's own teacher writes the training data, every model is scored on public benchmarks before it ships, and the same weights run in the browser, on the laptop and in the rack.

Why Palette

AI that Korean work can actually run on.

Tool calls on a phone, documents on a workstation, video in the cloud. One family, Korean-first, open where it can be, measured everywhere.