Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 9 November 2026. We have not tested K-Exaone 2.5 hands-on, and nothing here is a hands-on finding.
LG AI Research has opened its frontier-class AI model to the outside world — for free. The announcement came on Monday, 9 November 2026, at NVIDIA AI Day Seoul 2026, held at COEX in Seoul: the API for K-Exaone 2.5 is now available at no charge through LG's own platform, and anyone — developers, startups, companies, research institutions — can get an API key and start using the model for service development and research.
That is a genuine piece of news, and it is also a narrower one than the headline makes it sound. This article walks through what actually changed, what the model is, what the free API gives you, and where the honest limits sit.
What happened
LG AI Research announced on Monday that the application programming interface for K-Exaone 2.5, its frontier-class AI model, is open to the public free of charge. The API is available from today through the institute's own platform — the EXAONE API Platform at docs.exaone.ai — for what LG calls real-world validation.
The point of the free release, in LG's own words, is to broaden the chances for people to try a Korean-built AI model firsthand and integrate it into their own services. Lim Woo-hyung, co-director of LG AI Research, put it this way: "We are currently concentrating our capabilities on developing the third-generation model, and we hope that through this API release, more developers and companies will experience the improved performance firsthand. We will focus on completing a frontier-class Exaone model by January."
That last sentence matters, because it tells you what K-Exaone 2.5 is — and what it isn't.
What K-Exaone 2.5 is
K-Exaone 2.5 is a technology verification model. LG built it by applying additional training to the second-generation proprietary AI foundation model it released in August 2026. In plain language: it is a working, public step on the road to the third-generation model LG wants to finish by January — not the finished frontier model itself.
The techniques validated through 2.5 — spanning data design, post-training and reinforcement learning — will be folded into that third-generation model. So if you are trying it, you are testing a model that exists partly to prove the training approach for the next one. That is not a criticism; it is the design. But it is worth knowing before you build anything serious on top of it.
What we know about the model itself comes from LG's model card and platform documentation:
- 750 billion total parameters, 31 billion active — a mixture-of-experts design, so only a fraction of the model activates per token. As with every MoE model, the headline number is total capacity, not per-token compute.
- Text only, in Korean and English. No image, audio or video input.
- 512,000-token context window, with a combined input-and-output budget capped at 524,288 tokens.
- January 2026 knowledge cutoff.
- A
reasoning_effortparameter — low, high or max, with max as the default — that controls how much the model thinks before answering.
One more thing on the platform: it also lists EXAONE Tabular, a small model for predictions on structured data. The free API is not just the one big model.
What the free API actually gives you
The practical details are the best part of this story, because LG made the obvious developer-friendly choices:
-
The endpoint is OpenAI-compatible. If you already have client code that talks to an OpenAI-style chat completions endpoint, you point it at LG's platform with the model name
k-exaone_v2.5and it should just work. No new SDK to learn, no bespoke client to install. - It is free right now, with no price published. The platform's changelog says the service is currently free of charge. LG has not said how long the free period will last, or what limits will apply once it ends — so treat "free" as a launch condition, not a permanent price.
- You get an API key from LG's platform and use it for real services and research. This is not a demo playground with a toy quota for chat only; the stated purpose is service development and research.
That combination — a 750-billion-parameter model, a half-million-token context, an OpenAI-compatible endpoint, free — is genuinely worth trying if you build with LLMs. It is also the combination where the caveats matter most, because "free" from a lab that is still validating its training approach can change without much notice.
How to get an API key
Go to LG's EXAONE API Platform at docs.exaone.ai, sign up, and obtain an API key. Then point your existing OpenAI-compatible client at the platform's chat completions endpoint with the model name k-exaone_v2.5. The platform also documents the reasoning_effort parameter if you want to trade thinking time against latency. No pricing page to consult — there is no price yet — and no weights to download: access is through the platform only.
What you can build with it
The honest answer: anything you would build with a large text model that is strong in Korean and English, with a very large context window, and where you can tolerate a model that is still being validated.
- Agentic workflows and task automation. This is where LG is pitching the model hardest — its internal benchmarks focus on agentic tasks: automating work with SaaS tools, solving problems in workplace settings, finding information through web browsing, and software development tasks. The benchmark numbers are LG's own, from open datasets, and they are worth reading with that in mind (see below).
- Korean-language products. K-Exaone is built by a Korean lab as a Korean sovereign model. If you are building a service for Korean users, or handling Korean text at scale, a model designed around Korean from the start is a real option that most Western and Chinese models are not.
- Long-document work. A 512K-token context window is in the same class as the largest frontier models. Long contracts, long codebases, long research documents — the kind of work where a small context window forces you to chunk and stitch — is the natural use case, though, as ever, a big window on paper is not the same as reliable recall across all of it in practice.
- Research and evaluation. For researchers, a free frontier-class API is a chance to run evaluations and comparisons without a budget line. That is plainly part of why LG is doing this: more outside eyes on the model, more real-world validation data.
What you cannot do: run it locally. This release is API access only. LG has published weights for earlier K-Exaone models on Hugging Face, but no open weights have been announced for version 2.5. You use it through LG's platform or not at all.
The benchmarks — LG's numbers, labelled as such
LG says internal benchmarks on open datasets show K-Exaone 2.5 can compete with frontier-class global open-weight models in agentic AI. The comparison targets were DeepSeek V4.1 Flash and GLM 5.3 (max), China's leading frontier-class open-weight models. Here are the numbers LG reported, with the comparison models alongside:
- AutomationBench (automating tasks using SaaS tools): K-Exaone 2.5 scored 74.6, against DeepSeek 74.9 and GLM 71.6.
- Claw-Eval (an AI agent's problem-solving ability in workplace settings): 77.3, against DeepSeek 75.8 and GLM 77.6.
- BrowseComp (finding information through web browsing): 78.0, against DeepSeek 75.8 and GLM 79.0.
- SWE-Bench Pro (software development and modification tasks): 59.2, against DeepSeek 58.4 and GLM 64.0.
Read those numbers for what they are: a mixed picture, not a sweep. K-Exaone 2.5 sits roughly level with DeepSeek V4.1 Flash across the four tests and trails GLM 5.3 (max) on three of them. That is a credible result for a validation model — competitive with the field, not ahead of it — and it is LG's own internal testing, not an independent evaluation. No outside lab has published results on 2.5 yet, and we have not tested it ourselves.
The Nvidia connection
LG is also pursuing cooperation with Nvidia. It is reviewing the adoption of Nvidia's Dynamo technology to improve inference processing efficiency and service scalability, and Nvidia had earlier published a Dynamo usage guide tailored for K-Exaone. Dynamo is Nvidia's inference-serving framework, and the point of it here is practical: serving a 750-billion-parameter model to the public for free is expensive, and better inference efficiency is what makes a free API sustainable rather than a brief promotion.
The limits — stated plainly
- This is a technology verification model, not the finished frontier model. The full-performance model is the third generation, targeted for January. What you get today is the step that proves the training approach.
- The free period has no published end date — and no published limits. LG has not said how long free access will last or what will apply once it ends. Build on it, by all means, but do not bet a business on the price staying at zero.
- The benchmarks are LG's own internal results on open datasets. Competitive, roughly level with DeepSeek V4.1 Flash, trailing GLM 5.3 (max) on three of four tests. No independent evaluation exists yet.
- Text only, Korean and English only. No multimodal input, and if your product needs a third language as a first-class citizen, this is not the model for it.
- API only — no open weights for 2.5. Earlier K-Exaone models have published weights, but 2.5 is hosted access through LG's platform. You cannot self-host it, fine-tune it locally, or audit it independently.
- The January 2026 knowledge cutoff means the model does not know about anything after that date. For a model released in November 2026, that is a ten-month gap — fine for most coding and writing, worth remembering for anything current-events-adjacent.
- No hands-on testing was performed for this article. Everything above comes from the 9 November 2026 coverage and LG's platform documentation, checked the same day.
The bottom line
LG has put a genuinely large, genuinely capable model behind a free, OpenAI-compatible API — and the timing is not an accident. A Korean lab, on the stage at NVIDIA AI Day Seoul, offering the world a free look at its frontier work is a statement about where Korea wants to sit in the AI race: not as a customer of other countries' models, but as a builder of its own.
For developers, the sensible move is simple: get a key, point your existing client code at it, and run your own tests — especially if you work in Korean, or if you have agent workloads where a half-million-token context and a free price tag change the maths. Just do it with your eyes open: it is a validation model, the free window is undated, and the numbers so far are LG's own. Try it now, judge it on your own workload, and watch what January brings.
Sources: The Korea Herald, 9 November 2026, "LG AI Research opens K-Exaone 2.5 API for free" (https://biz.heraldcorp.com/article/10930227); Aju Press, 9 November 2026, "LG AI Research Opens Free Access to 'K-ExaOne 2.5' API" (https://www.ajupress.com/view/20261109100470731); Spectrum Web, 9 November 2026, "LG opens K-EXAONE 2.5 API free for developers" (https://www.spectrumweb.co/news/lg-k-exaone-2-5-free-api) — platform details, model card specs and endpoint reference via LG's EXAONE API Platform documentation (docs.exaone.ai). Benchmark figures are LG's internal results as reported by The Korea Herald. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.
Also published on ToolsDig: https://toolscout-e2i.pages.dev/posts/lg-k-exaone-2-5-free-api.html
Top comments (0)