EXAONE 3.0 7.8B

What EXAONE 3.0 7.8B is

EXAONE 3.0 is a large language model published by LG AI Research on 7 August 2024 ^[1]. It was the first model in the EXAONE family to be released openly, where earlier members of the family had stayed inside the company ^[1]. LG had built the family at several sizes; only one of them was made public, a 7.8B version ^[1].

A parameter is one of the numerical weights that training adjusts, and the count is a rough measure of a model's size rather than a direct measure of quality. "7.8B" means 7.8 billion such weights. The label matters mostly because it tells a reader what class of model this is and what hardware it implies.

Instruction-tuned, not a base checkpoint

The published model is instruction-tuned, not a base checkpoint ^[1]. The difference is worth spelling out. A base checkpoint is what comes out of pretraining on a large body of text: the model has learned to continue text plausibly, but it has no particular disposition to follow a request. Instruction tuning adds a further round of training on examples of instructions paired with good responses, which teaches the model to answer a question, follow a format, or take on a role. Releasing only the tuned version means users get something usable directly, while the pretrained starting point stays in-house.

Release and access

The weights were uploaded to Hugging Face on 31 July 2024 ^[2], slightly ahead of the 7 August publication date ^[1]. Access to the 7.8B weights is gated: a user must request access and be granted it before downloading ^[3]. Gating is a common arrangement for models whose publishers want a record of who is using them and an acceptance of terms. It does not make the model closed, but it does mean the download is not automatic: a request has to be approved first.

What LG emphasised

LG described the model as excelling in Korean while staying competitive on general tasks and complex reasoning ^[1]. That pairing — strength in one language alongside general competence — is the claim an open release has to support, because a model published with downloadable weights can be tested by anyone rather than taken on the strength of a paper alone.

Where it sits in the series

EXAONE 3.0 was followed in December 2024 by EXAONE 3.5, released in three sizes ^[2], so the line went from a single openly available model to a small family. It was extended again in March 2025 with EXAONE Deep, a set of reasoning models ^[2]. Reasoning models are built to work through intermediate steps before producing a final answer, spending more computation per response instead of emitting the first plausible continuation. The EXAONE line thus moved, within a fairly short span, from one gated instruction-tuned release to a range of sizes plus a reasoning variant.

The technical report sits on arXiv at version 4, last revised on 2 January 2026 ^[1], which indicates it was revised after its first posting rather than frozen at publication.