LG AI Research releases EXAONE 4.0 with a reasoning mode

What happened

On 15 July 2025, LG AI Research published EXAONE 4.0 ^[1]. The release marked a shift in LG's model line: from a pure language model to one with a switchable reasoning mode ^[1]. LG shipped two sizes at once, 32B and 1.2B ^[1]. The smaller 1.2B model is aimed at on-device deployment ^[1]. The 32B model had been uploaded to Hugging Face on 11 July 2025, four days before the report ^[2].

A switchable reasoning mode

A reasoning mode is a setting that lets a model work through intermediate steps before answering, while a non-reasoning mode aims at a direct response. EXAONE 4.0 was presented as a model that integrates the two modes behind a switch ^[1]. This is a common way to let one model serve both fast chat-style use and tasks that benefit from more deliberation. For users, the choice between a quick answer and a more deliberate one can sit inside one model family rather than requiring separate model lines.

Two sizes, two deployment targets

The pair of sizes covers different deployment needs. A 32B model is larger and generally needs more compute and memory; a 1.2B model is small enough to target local hardware. LG explicitly positions the 1.2B version for on-device deployment ^[1]. On-device deployment generally means running the model on the user's device rather than in a remote data center; this can matter for latency, connectivity, and local data handling. The 32B version is the one recorded on Hugging Face, uploaded on 11 July 2025 ^[2]. By shipping both sizes at launch, LG gave developers a choice between a compact on-device model and a larger model in the same release ^[1].

Claims and licence

LG described the models as better than open-weight models of the same class and competitive with frontier models ^[1]. That is the developer's positioning of the release. The release also changed licensing: commercial deployment through partners became allowed ^[2]. For organizations that want to build products on the model, the licence route matters alongside the model sizes and capabilities.

Why it matters

The significance is in the combination. EXAONE 4.0 offered a reasoning switch, two sizes, and a licence route for partners ^[1][2]. That made it relevant both to developers testing a larger model and to those looking at local, on-device use ^[1]. The 32B upload to Hugging Face gave the larger model a public location before the formal report ^[2]. And the licence change gave partners a path to commercial deployment ^[2]. The release therefore shows LG AI Research treating reasoning, deployment size, and licensing as parts of one product decision.