K-EXAONE 2.5: free API, 750B specs and benchmarks.
LG AI Research opened a free API for K-EXAONE 2.5, a 750B model (31B active) with 512K context. Specs, LG-reported benchmarks and how to try it.
By The Model Atlas · 2 sources
In this post7 sections
K-EXAONE 2.5 is LG AI Research’s newest large language model, and on 9 November 2026 the lab opened its API to everyone for free. LG announced it at NVIDIA AI Day Seoul 2026 in Seoul, and says anyone with an API key can use it at no cost for building services and for research. Here is what K-EXAONE 2.5 is, what LG says it can do, and how to try it.
What is K-EXAONE 2.5?
LG describes K-EXAONE 2.5 as a technical validation model: the second-round model from Korea’s Sovereign AI Foundation Model project, first shown in August, with additional training applied. LG says it used the model to test better training methods before building its third-round model.
According to the EXAONE API Platform documentation, K-EXAONE 2.5 is a text model for Korean and English with 750B total parameters and 31B active, a 512K-token context window, 512K maximum output tokens and a knowledge cutoff of January 2026. It supports chat, tool calling, structured outputs, code generation and streaming. The docs say it is especially strong at Korean understanding and generation, and that it handles long contexts and complex tool-calling instructions for business workflows and agentic tasks.
What’s new in K-EXAONE 2.5
LG says K-EXAONE 2.5 improved overall performance by more than 70% in eight weeks, based on its internal analysis across 24 evaluation metrics compared with the model released in August. LG credits its own training approach, covering data quality control, post-training and reinforcement learning. The docs add that a high-quality synthetic data method and an anomalous data detection system were used to curate the training data.
LG published four agentic benchmark results, measured by LG itself on open datasets. It compared the model with DeepSeek V4.1 Flash and GLM 5.3:
- AutomationBench (automating work with SaaS tools): K-EXAONE 2.5 scored 74.6, versus 74.9 for DeepSeek and 71.6 for GLM.
- Claw-Eval (agent problem-solving in work environments): 77.3, versus 75.8 for DeepSeek and 77.6 for GLM.
- BrowseComp (finding information by browsing the web): 78.0, versus 75.8 for DeepSeek and 79.0 for GLM.
- SWE-Bench Pro (software development and editing): 59.2, versus 58.4 for DeepSeek and 64.0 for GLM.
The docs also compare it with K-EXAONE 2.0: on AutomationBench it rose from 16.9 to 74.6, on BrowseComp from 50.87 to 78, and on SWE-Bench Pro from 47.1 to 59.21. LG’s own conclusion is that the model is competitive with frontier-class open-weight models on agentic tasks and “performing on par” with open-weight models of comparable scale. These are LG’s self-reported figures, so treat them as a starting point and test the model on your own work.
K-EXAONE 2.5 price and availability
The API is free. LG says that from 9 November anyone, including developers, startups, companies and research institutions, can get an API key from the EXAONE API Platform at docs.exaone.ai and use K-EXAONE 2.5 without paying. LG frames the free access as a way to let people try a Korean AI model in real services and to gather real-world usage. LG’s announcement and the model documentation we reviewed describe API access only and do not mention a weights download.
Who K-EXAONE 2.5 is for
It fits developers who want an agentic model for coding, search and tool use, and especially teams building Korean-language products. It is text only, so it is not an option if you need image input. Because LG calls it a technical validation model, it is also a preview of what LG plans next: the lab says it is now focused on its third-round model and aims to complete a frontier-level EXAONE model by January 2027.
How to try K-EXAONE 2.5
- Get an API key from the EXAONE API Platform (opens in a new tab).
- Call the model through the platform’s API using the K-EXAONE 2.5 model entry listed in the docs.
- Run your own coding, search and tool-calling tests next to the models you use today before switching anything in production.
How K-EXAONE 2.5 compares with other models on The Model Atlas
LG benchmarks K-EXAONE 2.5 against DeepSeek V4.1 Flash, which is already on The Model Atlas. For other large agentic models on The Model Atlas, see Kimi K3.1 Pro and LongCat-2.5, and browse what people are building with them. K-EXAONE 2.5 does not have a model page on The Model Atlas yet; one may be added after curation.
K-EXAONE 2.5 FAQ
What is K-EXAONE 2.5?
It is an upgraded version of LG AI Research’s K-EXAONE model, a text model with 750B total and 31B active parameters, according to LG’s API documentation.
Is K-EXAONE 2.5 free?
Yes. LG AI Research says its API has been free for everyone since 9 November 2026, after you get an API key.
What is the context length of K-EXAONE 2.5?
The EXAONE API Platform lists a 512K-token context window and 512K maximum output tokens.
Which languages does K-EXAONE 2.5 support?
Korean and English, with particular strength in Korean, according to LG’s documentation.
How does K-EXAONE 2.5 compare with DeepSeek V4.1 Flash and GLM 5.3?
In LG’s own tests it scored 74.6 on AutomationBench against 74.9 for DeepSeek and 71.6 for GLM, and 59.2 on SWE-Bench Pro against 58.4 and 64.0. LG says it is on par with open-weight models of comparable scale.
Can K-EXAONE 2.5 process images?
No. The documentation lists text as its only modality.
What comes after K-EXAONE 2.5?
LG says it is concentrating on its third-round model and aims to finish a frontier-level EXAONE model by January 2027.