LG opens free API access to K-EXAONE 2.5 for developers and researchers

LG AI Research opened free API access to K-EXAONE 2.5 on November 9, 2026, letting developers and researchers use the model in services without paying for access. The release pairs that invitation with LG’s claim of a substantial performance improvement from additional training, rather than a new generation of its foundation model.

An extra training round before the next generation

The announcement came at Nvidia AI Day Seoul 2026 at COEX in Seoul. K-EXAONE 2.5 builds on the second iteration of LG’s foundation model, released in August. LG calls the update a technology-validation model: it applies further training to that existing foundation ahead of work on a third iteration.

Kim I-reun, head of LG’s EXAONE Lab, said internal analysis found overall performance across 24 evaluation metrics improved by more than 70% against the August model. LG attributed the progress to advances in its training methodology, presenting the update as evidence that further gains were possible from the existing model.

Free access is intended to move that work into users’ hands. LG said it wanted people to try a Korean AI model directly and apply it to their services. Developers, startups, companies and research institutions can obtain an API key for service development and research.

The gains LG measured

Alongside the release, LG published four evaluations of agentic AI—systems that plan steps, use tools and work through tasks. The tests cover software changes, web research, workplace problem-solving and automation using online business software. K-EXAONE 2.5 improved over K-EXAONE 2.0 on all four in LG’s reported results.

Three reported gains from 2.0 to 2.5

SWE-Bench Pro

47.1→59.21

score

LG reported these scores for agentic coding, including software development and modification.

BrowseComp

50.87→78.0

score

LG’s reported web-search evaluation measures finding needed information online.

AutomationBench open-set

16.9→74.6

score

LG reported this improvement on automation tasks using software-as-a-service tools.

The competitor comparisons show a mixed picture rather than a sweep. On AutomationBench, LG’s 74.6 was just below DeepSeek V4.1 Flash at 74.9 and above GLM 5.3 (max) at 71.6. On BrowseComp, LG scored 78.0, between DeepSeek’s 75.8 and GLM’s 79.0.

For software development and modification, LG reported a SWE-Bench Pro score of 59.2, compared with DeepSeek’s 58.4 and GLM’s 64.0. These are LG’s internal measurements, not independent test results. LG lists code generation, debugging, refactoring and writing tests among the model’s intended applications.

What developers get—and still have to run

Hosted access is available through the EXAONE API Platform. As Unite.AI details, the model uses an OpenAI-compatible interface, including the POST /v1/chat/completions endpoint and model identifier k-exaone_v2.5. It supports chat, tool calling, structured outputs, code generation and streaming responses.

• Language and format: K-EXAONE 2.5 is text-only and supports Korean and English.
• Request size: input plus the maximum generation allowance cannot exceed 524,288 tokens, the units used to measure text. The allowance includes reasoning; larger requests return a 400 error.
• Reasoning controls: reasoning_effort defaults to max. LG’s guide warns that higher reasoning intensity can increase response time and token usage.

Tool calling does not mean the API carries out actions itself. In LG’s documented workflow, the model returns tool-call decisions. The surrounding agent software executes those calls, sends results back and manages subsequent turns. LG’s guide includes configurations for the Hermes general agent and OpenCode coding agent.

LG is also considering Nvidia’s Dynamo technology to improve inference efficiency—the work of generating responses—and service scalability. That remains a possible infrastructure step, not a confirmed deployment accompanying the free API release.