LG Opens K-EXAONE 2.5 API with 512K Context and Adjustable Reasoning Effort

On November 9, LG AI Research announced the opening of the K-EXAONE 2.5 API, allowing developers to test the model through a hosted service. The model focuses on Korean, English, and agentic tasks. The engineering focus of this update is how long context, tool calling, and reasoning controls can be integrated into existing applications. The official model page lists 750B total parameters and 31B active parameters per inference; both inputs and outputs are text-only. Announcement, model specifications

The API uses an OpenAI-compatible Chat Completions format, and its model identifier is k-exaone_v2.5. The context limit is 524,288 tokens, but the input length and max_tokens must sum to less than this limit; requests that exceed it return a 400 error. max_tokens also includes the reasoning process. Applications handling long documents therefore need to reserve room for reasoning and the final answer rather than filling the entire allowance with the prompt. API documentation

The model always reasons before answering, and thinking cannot be disabled. reasoning_effort accepts low, high, or max, with max as the default. Reasoning and the answer are placed in reasoning_content and content, respectively, and streaming is supported. Engineers integrating the model into agent frameworks need to ensure that reasoning fields, conversation history, and token accounting are handled correctly. The official API also offers an option to retain prior-turn reasoning for multi-turn tasks. Interface and parameter documentation

For capabilities, LG reports scores of 59.21 on SWE-Bench Pro and 78 on BrowseComp, compared with 47.1 and 50.87 for the previous generation. These figures indicate gains in the company’s internal evaluations, but they do not directly establish real-world project repair success rates, costs, or latency, nor do they constitute independent verification. Official evaluation table

NVIDIA confirmed that the model uses Nemotron datasets, including seven million synthetic persona records generated from Korean official statistics. For Chinese-speaking engineering teams, useful next steps include testing Traditional Chinese tool-parameter generation, information extraction from long documents, and latency differences across reasoning-effort settings. The main languages listed by the company are currently Korean and English, so Chinese capabilities require separate evaluation. NVIDIA technical announcement