LG AI Research has opened K-EXAONE 2.5 to developers through an application programming interface (API), and the service is “currently available free of charge”, according to the changelog of its EXAONE API Platform dated 9 November. LG lists the model at 750 billion parameters, with 31 billion active for each token.
On LG’s own benchmark table, K-EXAONE 2.5 scores 74.6 on AutomationBench, an agentic tool-use test, against 16.9 for K-EXAONE 2.0, about 4.4 times as high by our arithmetic. Nvidia’s blog says the model was “developed using NVIDIA Nemotron datasets”.
Business Pill · Sovereign AI Explained
An evergreen Business Engineer explainer on sovereign AI. It teaches the general idea only and does not discuss LG or this release.
The key insight: LG AI Research has opened K-EXAONE 2.5, a 750-billion-parameter model with 31 billion active, through an API that its changelog says is currently free of charge; on LG’s own table its AutomationBench score rose from 16.9 for K-EXAONE 2.0 to 74.6, about 4.4 times as high by our arithmetic.
The API is OpenAI-compatible, and K-EXAONE 2.5 “is a text-only model” with a maximum context length of 524,288 tokens, the endpoint documentation says. LG’s model page lists Korean and English as its languages, a knowledge cutoff of 2026-01, and chat, tool calling, structured outputs, code generation and streaming as features.
The platform launched on 9 November with two models: K-EXAONE 2.5, labelled “LLM(Sovereign AI)”, and EXAONE Tabular, a model for predictions on structured data. The pages we read do not say how long the free period will last.
Chosun Biz reported that LG AI Research made the announcement at Nvidia AI Day Seoul 2026 at COEX in Seoul, and that anyone can use the model “at no cost by simply getting an API key”. It described K-EXAONE 2.5 as “a technology validation model” that adds training to the second-generation model unveiled in August.
K-EXAONE 2.0 and 2.5 on the four agentic benchmarks LG AI Research published on its EXAONE API Platform Models page, 9 November 2026. The changes are our arithmetic; the scores are LG’s own.
How It Scores
LG’s table compares K-EXAONE 2.5 with its predecessor and with two Chinese open-weight models, DeepSeek-V4.1-Flash and GLM-5.3. The scores are LG’s own. On BrowseComp, a web-search test, it scores 78 against 50.87 for 2.0; on Claw-Eval, 77.29 against 62.11; and on SWE-Bench Pro, a software-engineering test, 59.21 against 47.1.
Against DeepSeek-V4.1-Flash, K-EXAONE 2.5 is higher on three of the four tests and lower on AutomationBench, 74.6 against 74.9. Against GLM-5.3, it is higher only on AutomationBench, 74.6 against 71.6, by our count of LG’s table. LG says the model performs “on par with open-weight models of comparable scale”.
Kim I-reun, head of the EXAONE Lab at LG AI Research, told Chosun Biz that “overall performance across 24 evaluation metrics improved by more than 70% compared with the model released in August”. LG has not published those 24 metrics on the pages we read.
What Changed From 2.0
K-EXAONE 2.0’s model card on Hugging Face lists the same 750 billion total parameters, with 37 billion active, a context length of 262,144 tokens and a 2025 second-quarter knowledge cutoff. LG released 2.0 with open weights under the Apache 2.0 licence.
Set against LG’s figures for 2.5, the context window doubles and the active parameter count is six billion lower, by our comparison of the two pages. LG’s documentation for 2.5 describes API access and does not mention open weights, and LG’s Hugging Face page listed no K-EXAONE 2.5 model when we checked on 9 November.
As we read LG’s pages, the release changes how the model is offered as much as how it scores: K-EXAONE 2.0 came with open weights, while the 2.5 documentation describes access only through LG’s own API.
As we read the benchmark table, the largest gain is on AutomationBench, LG’s tool-use test, while the coding gain on SWE-Bench Pro is smaller and GLM-5.3 still scores higher there.
As we read the timing, LG chose an Nvidia event in Seoul for the release, and Nvidia’s blog credits Nemotron datasets in the model’s training.
LG AI Research, EXAONE API Platform changelog, 9 November 2026
“This service is currently available free of charge.”
Three Implications
FOR DEVELOPERS LG’s endpoint documentation says the API is OpenAI-compatible and accepts up to 524,288 tokens of context.
FOR KOREAN AI POLICY LG’s platform lists K-EXAONE 2.5 under the label ‘Sovereign AI’.
FOR PRICING LG’s changelog says access is currently free; the pages we read give no end date or future price.
The framework says: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”
As we read it, LG has made access to K-EXAONE 2.5 free while its documentation describes the model only through LG’s own API, not as downloadable weights. The framework predates this release and does not discuss LG.
What Is Not Established
The benchmark scores are LG’s own and we did not test the model. LG has not said how long API access will stay free or what it will charge afterwards, and has not said whether K-EXAONE 2.5 weights will be released. The “24 evaluation metrics” behind the 70% figure are not listed on the pages we read. We did not contact LG AI Research.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
LG AI Research has opened K-EXAONE 2.5, a 750-billion-parameter model, through a free API, according to its platform’s changelog. On LG’s own table, its AutomationBench score rose from 16.9 to 74.6. LG says the model performs “on par with open-weight models of comparable scale”, naming DeepSeek-V4.1-Flash and GLM-5.3; on LG’s table, GLM-5.3 still scores higher on three of the four tests.
A note on sourcing. The specifications, benchmark scores and free access are from LG AI Research’s EXAONE API Platform pages, read on 9 November 2026; the Nemotron detail is from Nvidia’s blog; the event and the lab head’s quote are from Chosun Biz; the K-EXAONE 2.0 figures are from LG’s Hugging Face card. The multiples and comparisons are our arithmetic. We did not contact LG AI Research.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
