LG Opens Free API Access to K-EXAONE 2.5, Its 750-Billion-Parameter Reasoning Model
LG AI Research opened free API access to K-EXAONE 2.5, its 750-billion-parameter reasoning model, on 9 November 2026 at NVIDIA AI Day Seoul 2026 at COEX in Seoul, letting developers, startups, companies and research institutions obtain an API key through the EXAONE API Platform and use the model at no cost for service development and research. LG has not said how long the free access will last or what quota or rate limit will apply, so it should not be treated as free forever.
An API is a bridge that lets one program use the functions of another, in this case letting a developer send text to K-EXAONE 2.5 and receive an answer without running the model themselves. A reasoning model is a model that works through a problem step by step before answering, using extra internal thinking to handle coding, planning and multi-step tasks. K-EXAONE 2.5 is built for that kind of work, and LG positions it as an agentic model, meaning it can plan, use tools and complete real-world tasks.
The model is a technology-validation model, not a new generation. It was built by applying further training to K-EXAONE 2.0, the second proprietary foundation model LG released in August 2026 under the Korean government sovereign AI foundation model project, which aims to build homegrown AI models that reduce Korea reliance on foreign systems. The purpose of a technology-validation model is to test training techniques in the open and gather feedback, and LG plans to apply the techniques validated here to a third model, aiming to finish a frontier-level EXAONE model by January 2027.
The headline improvement is speed of progress, not just size. Kim I-reun, head of the EXAONE Lab at LG AI Research, said internal analysis confirmed that overall performance across 24 evaluation metrics improved by more than 70 percent compared with the model released in August, after only eight weeks of work on data quality and design, post-training and reinforcement learning. That figure is LG own internal claim, based on its own evaluations, and it has not been independently reproduced.
 |
What developers actually get
K-EXAONE 2.5 has 750 billion total parameters, with 31 billion active at any one time. The model uses a mixture-of-experts design, which means it has many specialist parts and activates only a subset for each task, keeping the active cost far below the total size.
The model handles text only, in Korean and English, with a January 2026 knowledge cutoff. It has a 512K token context window, with a combined input and output budget capped at 524,288 tokens, which is the same 512K limit expressed in exact tokens. That is a very large context, enough to hold long documents, large codebases or extended conversations in a single call. It supports chat, tool calling, structured outputs, code generation and streaming, and the endpoint is OpenAI-compatible, so existing client code can point at it with the model name k-exaone_v2.5.
Reasoning cannot be turned off in the API. A reasoning_effort parameter takes low, high or max, with max as the default, and controls how much the model thinks before answering. Reasoning tokens count towards the maximum token budget, so a long thinking process leaves less room for input and output.
Two models are listed on the EXAONE API Platform. The second is EXAONE Tabular, a small model for predictions on structured data. The platform changelog says EXAONE foundation models are now available through the API and that the service is currently free of charge. That wording is the key caveat: currently free does not mean permanently free, and LG has not published a free-tier quota, a rate limit or an end date. Developers who build a service on the free API should check the terms at docs.exaone.ai and plan for the possibility that pricing or limits appear later.
 |
How the benchmarks compare
LG published scores on four agentic AI benchmarks, comparing K-EXAONE 2.5 with China DeepSeek V4.1 Flash and GLM 5.3, and the results place it in the same range as those models without leading across the board. On SWE-Bench Pro, which tests software engineering, K-EXAONE 2.5 scored 59.21, against 58.41 for DeepSeek V4.1 Flash and 64.03 for GLM 5.3. On BrowseComp, it scored 78.0, against 75.8 and 79. On AutomationBench, which measures the ability to automate tasks, it scored 74.6, against 74.9 and 71.6. On Claw-Eval, which measures problem-solving in work environments, it scored 77.29, against 75.78 and 77.64.
Those numbers are vendor-run and not independently reproduced, so they should be read as LG own account of its model, not as a neutral ranking. The useful signal is not that K-EXAONE 2.5 wins every test, but that a Korean model built in eight weeks of further training now sits close to two well-known Chinese frontier models on agentic tasks. For Korean developers, the more important point may be language: the model is designed for Korean and English, and NVIDIA says it was trained with Nemotron datasets, including Nemotron-Personas-Korea, a set of 7 million synthetic Korean personas.
LG is also reviewing the use of NVIDIA Dynamo technology to improve inference efficiency and service scalability, and NVIDIA has already published a Dynamo usage guide tailored to K-EXAONE.
It is important to separate this release from an open-weights release. LG has published weights for earlier K-EXAONE models on Hugging Face, including the 750-billion-parameter K-EXAONE 2.0 under Apache 2.0, but it has not announced open weights for version 2.5. This release is API access only. Developers can call the model, but they cannot download its weights, fine-tune it locally, or run it on their own hardware. That is a deliberate choice, and it means the free API is a way to test the model, not a way to own it.
 |
| Illustration: Rows of server racks in a data centre, representing the infrastructure running K-EXAONE 2.5 |
What this means for Korean AI
For Korea, K-EXAONE 2.5 is part of a national push for sovereign AI, which means AI built on domestic models, data and infrastructure, tailored to Korean language, industries and priorities. The government sovereign AI project has backed several Korean models, and LG is one of the leading builders. A free API lowers the barrier for Korean startups and researchers to try a large domestic model without paying for foreign API access or building their own infrastructure.
For developers outside Korea, the model is most interesting as a Korean-language specialist. A model that handles Korean well, with a 524,288-token context, could be useful for Korean customer service, legal and financial documents, and research in Korean. The English support is also there, but the model does not list Chinese support.
| Specification | K-EXAONE 2.5 |
|---|---|
| Total parameters | 750 billion |
| Active parameters | 31 billion |
| Input type | Text only |
| Languages | Korean and English |
| Context limit | 524,288 tokens, 512K |
| Knowledge cutoff | January 2026 |
| API model name | k-exaone_v2.5 |
| Endpoint | OpenAI-compatible chat completions |
| Price at launch | Currently free of charge; no end date, quota or rate limit published |
| Benchmark, LG internal results | K-EXAONE 2.5 | DeepSeek V4.1 Flash | GLM 5.3 |
|---|---|---|---|
| SWE-Bench Pro | 59.21 | 58.41 | 64.03 |
| BrowseComp | 78.0 | 75.8 | 79 |
| AutomationBench | 74.6 | 74.9 | 71.6 |
| Claw-Eval | 77.29 | 75.78 | 77.64 |
Compare this model with the wider frontier race in our Carbon vs Opus 5.5 vs Astra One Point report, the hub of this cluster. The rival model picture is covered in our Google Gemini 4 Carbon is here article, the long-context open model in our Kimi K3.1 Pro is here report, and the open-weights side in our Mistral Large 4 Le Chonk is now open article.
Key Takeaways
- LG AI Research opened free API access to K-EXAONE 2.5 on 9 November 2026 at NVIDIA AI Day Seoul, via the EXAONE API Platform at docs.exaone.ai.
- The model has 750 billion total parameters with 31 billion active, handles text only in Korean and English, and has a 524,288-token limit, with an OpenAI-compatible endpoint.
- LG claims more than 70 percent improvement across 24 metrics in eight weeks, but that is an internal claim and has not been independently reproduced.
- On four agentic benchmarks, K-EXAONE 2.5 scored 59.21 on SWE-Bench Pro, 78.0 on BrowseComp, 74.6 on AutomationBench and 77.29 on Claw-Eval, close to DeepSeek V4.1 Flash and GLM 5.3.
- The service is currently free of charge, with no published end date, quota or rate limit, and no open weights have been announced for version 2.5.
Conclusion
K-EXAONE 2.5 is a fast, open test of Korean sovereign AI, not a finished frontier model. The free API gives developers a real chance to try a 750-billion-parameter Korean reasoning model, and the benchmark scores show it can sit beside DeepSeek and GLM on agentic tasks. Use it, test it in Korean, and plan for the free period to end, because LG has not promised that it will last.
Sources
Korea Times, LG AI Research opens K-EXAONE 2.5 API free of charge, 9 November 2026
Spectrum Web, LG opens K-EXAONE 2.5 API free for developers, 9 November 2026
AICoder, LG AI Research Opens Free K-EXAONE 2.5 API, 9 November 2026
EXAONE API Platform documentation, docs.exaone.ai
Comments
Post a Comment