Korea's Motif-3-Beta Ranks Top-3 Among Open Models

Korea's Motif Ships a 314B-Parameter Open Model It Says Ranks Third Globally Among Open Weights

A 30-person startup spun off from AMD-backed Moreh trained Motif-3-Beta on about 700 government GPUs — and became the highest-ranked model builder outside the US and China.

Jul 21, 2026Source: Hugging Face

South Korea has entered the frontier open-weights race. Motif Technologies, a roughly 30-person Seoul startup, released Motif-3-Beta — a 314-billion-parameter model whose weights hit Hugging Face around July 20, with the company's ranking announcement following in Korean press on Tuesday, July 21.

The architecture

Motif-3-Beta is a sparse mixture-of-experts design: 384 experts with 8 activated plus one shared per token, yielding about 13 billion active parameters — roughly 4% of the total — and a native 256K-token context window. The company says the architecture is fully in-house rather than a fork of existing open models. Training ran about five months on roughly 700 government-provided GPUs under Korea's sovereign-AI foundation model program, which selected Motif alongside LG AI Research, SK Telecom and Upstage.

The ranking claim

Per the company, citing Artificial Analysis scores, Motif-3-Beta's 44 on the Intelligence Index makes it third among open-weights models globally — behind Moonshot's Kimi K3 and Zhipu's GLM-5.2, tied with DeepSeek V4 Pro — and 11th among all models, the best showing for any lab outside the US and China. The framing is Motif's own reading of Artificial Analysis data, and the benchmarked checkpoint predates the public weights.

The fine print

The beta's license allows personal, educational and research use only; commercial use requires written permission — so 'open-weights, non-commercial' is the accurate label for now. A final version with improved performance is promised for early August, and the government program targets full commercial open-source for the finished model.

The AMD lineage

Motif spun off from Moreh, the Korean AI-software company whose 2023 Series B included AMD and KT — and whose stack runs on AMD Instinct GPUs. Motif raised its own ₩24 billion (~$16.6 million) Series B in May. A frontier-adjacent model built outside the Nvidia ecosystem, by a team of 30, is precisely the data point the open-weight cost debate did not need this week.

Why It Matters

The open-weight frontier is no longer a US-China duopoly: a 30-person Korean team with 700 GPUs just landed within striking distance of models trained on hundreds of thousands. That collapses the assumed capital floor for sovereign AI — every mid-size country now has an existence proof — and gives AMD a rare frontier-model showcase outside Nvidia's stack.