LG opens K-EXAONE 2.5 API for free with noncommercial-use terms

LG AI Research opened free API access to K-EXAONE 2.5 on November 9, allowing developers to call its hosted model for Korean and English responses and code generation. LG announced the opening at NVIDIA AI Day Seoul 2026.

The platform’s terms grant noncommercial use. They prohibit using the model or its outputs for profit and connecting the API to a product or service so that third parties can call the model or submit inputs. LG’s announcement welcomes developers, companies and research institutions; access remains subject to those platform restrictions. Users must be at least 19.

The model documentation lists 750 billion total parameters, with 31 billion active, and text support in Korean and English. Capabilities include conversation, tool calling, structured responses and code generation. Its listed knowledge cutoff is January 2026.

The endpoint allows up to 524,288 tokens across the input and generation allowance together. Tokens are the small units a model processes; the space reserved for generation includes reasoning as well as the answer. An OpenAI-compatible chat interface offers complete responses or streamed output.

For coding-agent connections, the surrounding agent program executes the tool calls returned by the model. LG’s integration guide describes that program passing execution results back to the model and continuing through subsequent requests to the final answer. Model decisions, computer actions and the follow-up response form successive parts of the workflow. With streamed tool calls, the guide calls for checking that arguments are complete before execution. It says higher reasoning intensity can increase response time and token use, and treats that setting separately from carrying earlier reasoning and tool-call history into subsequent requests.

Requests authenticate with a secret API key sent in an authorization header. LG’s authentication documentation says keys belong on the server, kept out of browser or app client code. The platform terms give each issued key a 30-day validity period.

Consent determines storage of submitted content

Retention of prompts, outputs and request metadata for training and evaluation requires optional consent, withdrawable in the profile. Retention lasts until one year after account withdrawal or consent withdrawal. Declining users’ content is discarded after processing; material already trained into models cannot be separated and deleted.

Request limits also apply. The rate-limit documentation describes per-second, per-minute and daily windows, remaining-request headers and a 429 error when a limit is exceeded. Figures such as one request per second and 60 per minute appear in an example response, not as a guaranteed allocation for every account. Headers indicating how long to wait before another request may be absent.

In LG’s own published table, the SWE-Bench Pro software-modification score rises from 47.1 for K-EXAONE 2.0 to 59.21 for 2.5. The comparison entries are 58.41 for DeepSeek V4.1 Flash and 64.03 for GLM 5.3. LG measured these results using open datasets; they are benchmark scores, not direct measurements of production response speed or developer-team productivity.

LG describes 2.5 as a technology-validation model, further trained from the second version released in August. It plans to incorporate the validated training techniques into a third model now under development. Co-head of LG AI Research Lim Woo-hyung set a goal of completing a frontier-performing EXAONE model by January 2027.