Samsung drops Gauss2 AI model in three distinct sizes

Most tech giants introduce a new flagship AI model by putting up benchmark charts and picking fights with rivals. Samsung is taking a very different route: testing it on thousands of its own employees first.

At the Samsung Developer Conference Korea 2024, the company officially unveiled Gauss2, the second generation of its in-house generative AI model.

Three sizes, from on-device to MoE

Samsung split Gauss2 into three distinct flavors, each targeting a specific compute environment.

First up is the Compact model. It is optimized strictly for on-device processing on hardware with limited compute, meaning it handles tasks locally without calling home to a server.

The Balanced model targets the middle ground, aiming for the sweet spot between execution speed, general performance, and resource efficiency.

At the top sits the Supreme model. It leans on a Mixture of Experts (MoE) architecture, activating specialized sub-networks to keep cloud inference costs down without tanking performance.

Gauss2 supports between 9 and 14 natural languages alongside various coding languages. Samsung built its own custom tokenizer and stabilization techniques specifically to sharpen its coding and multilingual translation chops.

A company-wide sandbox

Paul Kyungwhoon Cheun, President and CTO of Samsung’s DX Division, made it clear that Gauss2 is already doing heavy lifting behind closed doors.

Take code.i, Samsung’s internal coding assistant. Since rolling out last December, monthly usage has surged fourfold. Roughly 60% of software developers in Samsung's Korean DX division now actively code with it.

There is also the Samsung Gauss Portal, a conversational assistant that drafts emails, translates documents, and summarizes briefings for corporate employees across global subsidiaries. Even customer support call centers are using Gauss-powered tooling to categorize and summarize service logs automatically.

Next, Samsung wants to bolt on multimodal skills, like analyzing charts and generating images, alongside knowledge graph tech under its broader "AI for All" push.

Why it matters

While you cannot go to a website and chat with Gauss2 today, you will almost certainly interact with it through your next Samsung device. The Compact model is clearly engineered for future Galaxy phones and tablets, promising faster, private, on-device execution for everyday tasks.

Strategically, we think Samsung is playing this smart. Instead of torching billions trying to out-reason frontier models in the cloud, it focused on MoE efficiency and on-device constraints. That plays directly to its superpower: shipping hundreds of millions of consumer hardware units every year.

The real test will be execution. Packaging capable models onto mobile thermal envelopes without gutting battery life is notoriously difficult. But by treating 60% of its engineering workforce as guinea pigs for a year, Samsung at least enters the consumer arena with battle-tested software.