LG AI Research introduced K-EXAONE 2.5 on November 9, 2026, and opened its API free of charge. According to LG, anyone who obtains an API key on the EXAONE API Platform can use the model at no cost for development and research – developers, startups, enterprises and research institutions alike. The launch took place at NVIDIA AI Day Seoul 2026. The benchmark figures LG is promoting, however, come exclusively from internal measurements. Here is a sober look at what is known and what is not.
What K-EXAONE 2.5 is – and what it isn’t
LG itself calls K-EXAONE 2.5 a “technical validation model”. It is not the next big model generation but an additionally trained version of its predecessor K-EXAONE 2.0, which LG developed as part of South Korea’s programme for independent AI foundation models. According to LG’s Korean press release, the team wanted to test how far its existing training methodology could be pushed before starting work on the third model. The findings are meant to feed into that third-round model.
The technical specifications are listed in the model overview of the EXAONE API Platform:
- Parameters: 750 billion in total, 31 billion active (mixture-of-experts design)
- Context window: 512K, exactly 524,288 tokens according to the API documentation; the same figure is given as the maximum output
- Modality: text only
- Languages: Korean and English
- Knowledge cutoff: January 2026
- Features: chat, tool calling, structured outputs, code generation, streaming
On training, LG only says in general terms that a methodology based on high-quality synthetic data was applied from continual pre-training through post-training, and that a system for detecting anomalous training data was used. There are no specifics on data volume, training budget, or a model card with safety evaluations.
Important for anyone who wants to self-host: there are no weights for K-EXAONE 2.5 so far. On Hugging Face, the LGAI-EXAONE organization still lists the K-EXAONE 2.0 repositories from July 2026 as its newest entries as of November 10. For now, K-EXAONE 2.5 is an API-only model.
The benchmarks: vendor figures with a chosen comparison
LG highlights four benchmarks meant to measure agentic capabilities – tasks in which a model plans on its own, uses tools and works through several steps. The platform documentation gives the following scores (all according to LG):
- SWE-Bench Pro (software development): K-EXAONE 2.5 59.21 – K-EXAONE 2.0 47.1 – DeepSeek-V4.1-Flash 58.41 – GLM-5.3 64.03
- BrowseComp (web search): 78 – 50.87 – 75.8 – 79
- AutomationBench, open set (automation with SaaS tools): 74.6 – 16.9 – 74.9 – 71.6
- Claw-Eval (problem solving in a workplace setting): 77.29 – 62.11 – 75.78 – 77.64
In the English announcement by LG AI Research, the scores are rounded to 59.2, 78.0, 74.6 and 77.3. The wording there is that the model delivers performance “comparable” to leading open-weight models and is “competitive” with DeepSeek-V4.1-Flash and GLM-5.3.
On closer inspection, that is a fair summary – but no more than that. K-EXAONE 2.5 is narrowly ahead of DeepSeek-V4.1-Flash in two of the four disciplines and narrowly behind on AutomationBench. Against GLM-5.3, it only leads on AutomationBench; on SWE-Bench Pro, GLM is ahead by roughly five points. This is catching up, not outperforming.
What stands out most is the jump over its own predecessor: on AutomationBench, the score rises from 16.9 to 74.6, on BrowseComp from 50.87 to 78. In the release, Kim Yi-reun, head of the EXAONE Lab, is quoted as saying that overall performance across 24 benchmarks improved by more than 70 percent over the model released in August – after eight weeks of work. LG does not disclose which 24 benchmarks these are or how the overall figure is calculated.
For context: all scores come from internal measurements on public datasets. There are no independent reproductions yet. With agentic benchmarks, results also depend heavily on the agent harness, tool definitions and time limits – details LG does not document in the sources we reviewed.
The API in practice
Getting started is straightforward. According to the documentation, the EXAONE API is OpenAI-compatible; requests go to the POST /v1/chat/completions endpoint with the model ID k-exaone_v2.5. Anyone using existing clients or agent frameworks that speak the OpenAI format mainly needs to swap the base URL, key and model name.
One peculiarity: K-EXAONE 2.5 always “thinks” before it answers. Reasoning cannot be switched off; it can only be tuned via the reasoning_effort parameter with the levels low, high and max. The default is max. LG recommends low for fast responses and high for maths and coding. The reasoning is returned separately in reasoning_content, and the tokens it consumes are reported in usage.reasoning_tokens. For multi-turn agents, the documentation recommends keeping earlier reasoning via chat_template_kwargs.preserve_thinking.
Anyone testing the model should keep in mind that, according to the documentation, max_tokens includes the reasoning. A tight limit can therefore leave little room for the actual answer.
Free – but with strings attached
On pricing, the situation is simple: there is none. LG does not name an end date for the free phase in the sources we reviewed. The release speaks of opening the API “to validate it in real-world use”. Put differently, LG wants to see usage data.
A detail from the rate-limit documentation fits that picture. Limits per second, minute and day are reported in HTTP headers. Users who opt in to “Service Improvement Analytics” in their profile get a higher daily quota. LG says consent is optional and the API works without it. For European companies processing personal data, it is still something to review against data-protection and data-processing requirements before deployment.
Analysis: catching up with a clear agenda
The context is political. K-EXAONE is part of South Korea’s push to build its own foundation models and reduce reliance on foreign systems. In the release, LG co-head Lim Woo-hyung says the team is focusing on the third-round model and aims to complete an EXAONE model with frontier-level performance by January. LG also says it is reviewing NVIDIA’s Dynamo inference software to improve efficiency.
The fact that LG benchmarks itself against Chinese open-weight models is both honest and revealing. Models such as DeepSeek V4.1-Flash and GLM 5.3 are currently the yardstick in the open-weight space, as is Moonshot’s Kimi K3.1 Pro. The difference: their weights are available, K-EXAONE 2.5’s are not. An API model that is merely “on par” with freely available models has to win developers over on other factors – price, availability, Korean-language quality or trust in the provider.
Right now, the price is unbeatable. For teams building Korean-language applications, the model may be a serious candidate, especially since LG emphasizes its strength in Korean. For German- or other European-language use cases, K-EXAONE 2.5 is of limited relevance: only Korean and English are officially supported.
What is still missing
- Independent benchmarks: so far, there are only LG’s numbers.
- Model card and safety report: the sources we reviewed contain no information on safety evaluation or red-teaming.
- Pricing after the free phase: neither the duration nor later terms are public.
- Weights: whether K-EXAONE 2.5 will appear on Hugging Face like its predecessor remains open.
Bottom line: K-EXAONE 2.5 is above all a progress report – a model that has caught up considerably, with a large context window and a free API, whose scores still need to be confirmed outside LG. Anyone testing agentic workflows can run it against their own tasks at no cost – and should trust their own measurements more than the table from Seoul.
