At NVIDIA AI Day Seoul 2026 in Seoul, LG AI Research announced free API access to K-Exaone 2.5, inviting developers, companies, startups and researchers to test the upgraded model on real-world tasks.

The text-only model has 750 billion total parameters, with 31 billion active per token, and supports Korean and English, tool calling, structured outputs, code generation and streaming through an OpenAI-compatible API; its context window reaches 524,288 tokens.

LG highlights coding tasks such as generation, debugging, refactoring and test writing, as well as agentic web search and workplace automation.

In LG’s internal evaluations, K-EXAONE 2.5 scored 59.21 on SWE-Bench Pro, 78.0 on BrowseComp, 74.6 on AutomationBench open-set and 77.29 on Claw-Eval, improving on its predecessor across all four tests.

Its results were competitive with the cited Chinese models DeepSeek V4.1 Flash and GLM 5.3 (Max), but varied by benchmark: K-EXAONE 2.5 beat DeepSeek on three of four tests and GLM on one.

LG says the model improved by more than 70%, but that figure comes from internal aggregate evaluations whose baselines, metric weighting and task details were not disclosed; coverage also differs on whether the comparison point was July or August.

The benchmark results have not been independently tested, and NVIDIA’s blog says the model was developed using NVIDIA Nemotron datasets.

Developers can adjust reasoning with reasoning_effort, which defaults to max, and preserve_thinking, which carries prior assistant reasoning into later turns; LG cautions that more reasoning can increase latency and token use.

For agent workflows, the surrounding harness—not the model API—executes tool calls, returns their results in later turns and manages the interaction loop; LG provides setup guides for Hermes and OpenCode.

LG has not specified API limits, access duration, latency targets, serving hardware, data-retention terms or uptime commitments, and it has not announced downloadable K-EXAONE 2.5 weights.

Free access gives small teams and researchers a way to prototype Korean-language services without usage fees, but unclear quotas, data terms, reliability and future pricing make production reliance difficult to assess; developers are advised to test real workloads, protect sensitive data and API keys, and keep a fallback model.

LG plans to apply verified post-training techniques to its third-phase foundation model and aims for a globally frontier-class EXAONE model by January 2027, while exploring NVIDIA Dynamo to improve inference efficiency and service scalability.