LG AI Research Releases K-EXAONE 2.0, a 750B-Parameter Open-Weight MoE Model with 262K Context

LG AI Research Ships 750B-Parameter K-EXAONE 2.0

LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter mixture-of-experts (MoE) language model with 37 billion active parameters per forward pass, under the Apache 2.0 license. The model is available now on Hugging Face (LGAI-EXAONE/K-EXAONE-2.0-750B-A37B).

According to LG AI Research, the model was built by "upcycling" its predecessor, K-EXAONE (236B total / 23B active parameters), expanding it to more than three times the size through depth and width scaling, followed by continual pretraining, difficulty-focused mid-training, and post-training.

Architecture and Specs

• Total parameters: 750B (37B active)
• Layers: 78 (2 dense + 76 sparse), plus 1 MTP layer
• Attention: 64 heads (64 Q-heads / 8 KV-heads), head dimension 128, mixing global (NoPE) and sliding-window attention (SWA)
• Experts: 256 total, 8 activated, 1 shared expert, expert dimension 2,048
• Vocabulary: 153,600 tokens
• Context window: 262,144 tokens
• Knowledge cutoff: Q2 2025
• License: Apache 2.0

LG AI Research says a modification to the SwiGLU activation function — clamping after the two SwiGLU branches — mitigated exploding activations in deeper layers, improving training and inference stability at this scale.

Benchmark Claims

LG AI Research positions K-EXAONE 2.0 against Qwen3.5, GLM-5.1, and DeepSeek-V4 Pro (max) on a battery of internal evaluations. Reported scores include:

• MMLU-Pro: 83.5 (vs. 89.8 Qwen3.5, 86.0 GLM-5.1, 87.5 DSV4 Pro)
• GPQA-Diamond: 82.2 (vs. 88.4 / 86.2 / 90.1)
• AIME 2026: 92.3 (vs. 91.3 / 95.3 / 95.2)
• SWE-Bench Verified: 68.2 (vs. 76.4 / 73.6 / 80.6)
• OpenAI-MRCR (long context): 94.4 (vs. 93.0 / 71.5 / 92.9)
• KGC-Safety: 99.8 (vs. 92.0 / 69.3 / 82.8)

These figures come directly from LG AI Research's own technical report and have not been independently verified. The company claims particular strength in long-context retrieval and safety, where K-EXAONE 2.0 leads or closely trails larger competing models despite having fewer active parameters than GLM-5.1 (40B) and DeepSeek-V4 Pro (49B).

Multilingual Expansion and Inference

Language coverage expanded from six to ten: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, and Portuguese.

For inference, LG AI Research says it supports two speculative decoding methods — Multi-Token Prediction (MTP) and DSpark — claiming 3–5x generation speedups for latency-sensitive, long-horizon agentic workloads. The company provides forked SGLang and vLLM installations for serving the model, with reference configurations for two-node clusters of 8x NVIDIA H200 GPUs. DSpark is not yet supported on vLLM.

Pricing is not applicable in the traditional sense — this is an open-weight release requiring self-hosted infrastructure. No hosted API pricing has been disclosed.

What This Means

K-EXAONE 2.0 puts LG AI Research into the small group of labs releasing frontier-scale (700B+ parameter) open-weight models, joining the ranks of DeepSeek and Zhipu AI's GLM line. The MoE design with only 37B active parameters keeps inference costs manageable relative to the total parameter count, though running it still requires multi-node H200 clusters — this is not a model for single-GPU deployment.

The benchmark story is mixed: K-EXAONE 2.0 trails Qwen3.5, GLM-5.1, and DeepSeek-V4 Pro on most reasoning and coding tasks, but claims meaningful leads in long-context retrieval (OpenAI-MRCR) and safety evaluations (KGC-Safety, ROK-Fortress). For teams prioritizing Korean-language performance, long-context reliability, or safety compliance over raw benchmark leadership, this could be a relevant option — provided the infrastructure to run a 750B-parameter model is available. As with all self-reported benchmarks, independent verification will determine whether these claims hold up in practice.