LG opens K-EXAONE 2.5 API free after 70% overall jump

By Sophie Lindqvist

LG AI Research, the artificial intelligence research arm of LG Group, opened the K-EXAONE 2.5 API at no cost on Monday after reporting a performance gain of more than 70% across 24 benchmarks in eight weeks. The company announced the launch at NVIDIA AI Day Seoul 2026, held at Coex in southern Seoul.

Starting that day, LG said developers, startups, companies and research institutes could use K-EXAONE 2.5 in services and research without paying. They can obtain an API key through the company’s platform at docs.exaone.ai.

The release sits inside South Korea’s government sovereign AI foundation model project. LG said K-EXAONE 2.0 was the second model it developed under the program, which aims to build homegrown AI models and reduce Korea’s reliance on foreign systems.

What exactly did LG release?

LG AI Research describes K-EXAONE 2.5 as a technology verification version of its upgraded foundation model. It built the model by applying additional training to K-EXAONE 2.0. Its new API makes that model available for outside use at no charge.

The company also linked the release to a planned development sequence. LG will apply training techniques validated through K-EXAONE 2.5 to a third-round model. The verification label makes that relationship explicit: this release tests a set of methods before LG carries them forward.

LG put the model in front of outside users, not just its own researchers. Developers, startups, companies and research institutes can request keys and use it in working services or research. That gives them direct access to the verification build that sits between K-EXAONE 2.0 and LG’s next model.

For users, the immediate offer is a no-cost API key for a model LG still treats as a technology test.

How strong is K-EXAONE 2.5 on agentic AI?

LG AI Research’s own tests on open datasets put K-EXAONE 2.5 on par with leading Chinese open-weight models DeepSeek V4.1 Flash and GLM 5.3 (max) in agentic AI. LG describes agentic AI as systems that plan tasks, use tools on their own and complete real work.

AutomationBench provided the clearest direct comparison. The benchmark measures task automation using software-as-a-service tools.

• AutomationBench: K-EXAONE 2.5 scored 74.6, compared with 74.9 for DeepSeek V4.1 Flash and 71.6 for GLM 5.3 (max).
• Claw-Eval: K-EXAONE 2.5 scored 77.3 on a test of workplace problem-solving.
• BrowseComp: K-EXAONE 2.5 scored 78.0 on a test of web browsing.
• SWE-Bench Pro: K-EXAONE 2.5 scored 59.2 on a test of software development.

The AutomationBench result placed K-EXAONE 2.5 close to DeepSeek V4.1 Flash and ahead of GLM 5.3 (max), based on LG’s reported scores. LG characterized its broader agentic performance as on par with both models, although the announcement supplied direct competitor figures only for AutomationBench.

That broader claim rests on an internal analysis covering 24 benchmarks.

“Internal analysis showed that overall performance across 24 benchmarks improved by more than 70 percent compared with the model released in July,” said Kim Yi-reun, head of the EXAONE Lab.

The more-than-70% gain covers aggregate performance across the 24 tests, rather than a score on any single benchmark. It compares K-EXAONE 2.5 with the model LG released in July. LG described the improvement as arriving in just eight weeks.

Because these results came from LG’s own tests on open datasets, they document the company’s claim rather than independent validation. On LG’s evidence, K-EXAONE 2.5 sits close to two leading open-weight rivals on agentic tasks, subject to independent testing.

Why does the free API fit Korea’s sovereign AI goal?

LG’s sovereign AI work extends beyond K-EXAONE 2.0. The company developed 2.0 as the second model under the government’s sovereign AI foundation model project. The stated policy aim is to build Korean AI models that reduce the country’s reliance on foreign systems.

K-EXAONE 2.5 extends that line through further training. The free API then places the project in front of four groups named by LG: developers, startups, companies and research institutes. Users can test it without a charge and build it into services, according to the announcement.

This gives the release two practical dimensions. LG is testing training methods on a model it plans to refine, while outside users gain access to the same model at no direct cost. The policy objective and the technical test point in the same direction.

The model also has a named competitive frame. LG measured it against DeepSeek V4.1 Flash and GLM 5.3 (max), which it described as leading Chinese open-weight models. Under the project’s stated aims, a Korean model that approaches those results would support the goal of reducing dependence on foreign systems.

The no-cost API lowers the barrier to trying LG’s homegrown model while keeping the government-backed sovereignty goal in view.

What could NVIDIA’s Dynamo change?

LG AI Research said it is also reviewing NVIDIA’s Dynamo technology to improve inference efficiency. Its announcement describes an option under evaluation, not a deployed integration.

Capability and serving efficiency address different parts of putting a model into production. LG’s benchmark scores cover tool use, web browsing and software development. Dynamo concerns the efficiency of running the model to produce those results.

By reviewing Dynamo, LG is considering NVIDIA’s serving technology as a route to better inference efficiency. The announcement did not set a deployment date or name a product schedule for that work.

Dynamo remains under review, while the free API offer is already available.

What is next for the third-round EXAONE model?

LG plans to apply the training techniques validated through K-EXAONE 2.5 to its third-round model. That makes the free API part of a broader development sequence: outside access to a verification model and a planned transfer of its methods into the next release.

“We will focus on completing an EXAONE model with frontier-level performance by January next year,” said Lim Woo-hyung, co-head of LG AI Research.

That target follows the reported 70% aggregate improvement and the agentic comparisons with DeepSeek V4.1 Flash and GLM 5.3 (max). LG does not present K-EXAONE 2.5 as its finished frontier model. It presents the release as a technology verification step toward a model due by its stated deadline.

For outside users, the current benchmark scores establish a baseline against which a later EXAONE release can be compared. LG has also set a deadline for applying the methods that produced this gain.

The next checkpoint is LG’s third-round EXAONE model, which the company says will apply K-EXAONE 2.5’s validated training techniques and target frontier-level performance by January next year.