EXAONE 3.5

What EXAONE 3.5 is

EXAONE 3.5 is a family of large language models from LG AI Research. All three members are instruction-tuned, meaning they are trained not only to continue text but to carry out tasks given as prompts. The series comes in three sizes — 2.4B, 7.8B and 32B parameters — where the number describes the scale of the model's learned weights and, roughly, how much computation each generated token costs^[1][2]. All three are bilingual, working in both English and Korean^[2].

The series succeeds EXAONE 3.0, the 7.8B instruction-tuned model LG published in August 2024^[3].

How it was released

EXAONE 3.5 was announced on 6 December 2024 together with a technical report describing the models and their evaluation^[1]. The weights, however, were already on Hugging Face on 1 December 2024, five days before the report appeared^[2]. The downloadable models therefore preceded the write-up that documents them.

Sizes, architecture and context

The smallest model, 2.4B, is aimed at constrained devices, where memory and compute are limited^[1][2]. The 7.8B model's configuration is documented in more detail: 6.98 billion parameters excluding embeddings, 32 layers, and a vocabulary of 102,400 tokens^[2]. Its attention uses grouped-query attention with 32 query heads and 8 key-value heads^[2].

Grouped-query attention is a way of saving memory when a model runs. In a standard design each query head has its own key and value projections, and for long inputs the stored keys and values become a large part of the cost. Grouped-query attention lets several query heads share one set of key and value projections, so that store shrinks. Here 32 query heads share 8 key-value sets, four query heads to each.

All three sizes accept context of up to 32K tokens^[2]. Context is the text the model can attend to in a single pass — the prompt together with whatever material it has been asked to work over.

What the report claims

The technical report states that the series scored highest across seven instruction-following benchmarks, and also reports top performance in four long-context benchmarks^[1]. Those are the authors' comparisons against other models; the benchmark names and the margins are set out in the report itself. Instruction-following benchmarks test whether a model actually does what a prompt asks rather than producing something merely plausible, while long-context benchmarks test whether it can find and use information scattered through a long input.

Licence and where it sits

Weights are released under the EXAONE licence, which permits research use^[1][2]. The family is organised as a ladder rather than a single model: a small one for devices with tight memory, a mid-size one whose architecture is published in detail, and a 32B one for larger machines, all sharing the same 32K-token context and the same English–Korean pairing^[1][2]. For readers studying how model families are built and documented, the package is unusually complete — weights, a model card with layer and attention details, and a technical report describing the evaluation, with the gap between the weight release and the report showing how the two halves of a release can arrive on different days^[1][2].