Reported 9 Nov 2026 · Labs A one-minute read

LG opened Korea's state-backed national AI model to any developer, free.

LG AI Research said on November 9 that the K-EXAONE 2.5 interface is now free to developers, companies and research institutions. Caveat: the weights stay closed, and no end date for free access was given.

Until now

A government-backed national model is rationed to approved partners and industry.

Now

Korea's is free to anyone, priced below every foreign model it competes with.

Why it matters

Free access is not generosity, it is data collection and standard setting. LG says the point is testing in real services. A national model that becomes the cheapest way for Korean developers to build is the one Korean software is written against, which is what sovereign AI means in practice.

Established

LG AI Research announced free interface access on November 9 in Seoul.
Access needs a key from LG's own platform; no price is published.
LG calls 2.5 a technology verification model, not the finished generation.

Not established

That the benchmark scores hold. They are LG's own internal evaluation.
How long free lasts, or what limits follow. Neither was stated.
That weights will be released. No open-weight plan was announced.

The five-minute readThe record, the timeline, the weight and 6 sources

The record

Reported, not verified: we could not reach an LG primary. docs.exaone.ai carries no changelog or pricing page we could open, and we found no release on LG AI Research's own site. We read 4 Korean outlets, 2 of them timestamped, which agree on the date, venue and terms. Reading LG AI Research's own release, or a dated free-access entry on the EXAONE platform, moves this card to verified by November 20.

LG AI Research announced on Monday November 9, 2026 at NVIDIA AI Day Seoul 2026, held at COEX in Seoul's Gangnam district, that the K-EXAONE 2.5 interface is open to everyone at no charge. Seoul Economic Daily published at 10:00 Korea Standard Time, which is 01:00 Universal Time, and DigitalToday at 10:18, updated at 14:00. Both say developers, startups, companies and research institutions can take an interface key from LG AI Research's own platform and use the model for service development and research at no additional cost. LG describes 2.5 as a technology verification model: additional post-training applied to K-EXAONE 2.0, built in the second phase of its proprietary foundation model project. The K-EXAONE series sits inside the Ministry of Science and ICT's programme to build an independent national foundation model. Kim Yi-reun, head of the EXAONE Lab, is quoted saying internal analysis confirmed performance across 24 combined metrics improved more than 70% against August's model, and that core training technology was secured in-house. Lim Woo-hyung, head of LG AI Research, is quoted on the third phase: a model with frontier-class performance by January. K-EXAONE 3.0 is in development, planned for completion in January 2027. The benchmark table reported, all LG's own runs on public datasets, gives K-EXAONE 2.5 against DeepSeek V4.1 Flash and GLM 5.3: AutomationBench 74.6 against 74.9 and 71.6; Claw-Eval 77.3 against 75.8 and 77.6; BrowseComp 78.0 against 75.8 and 79.0; SWE-Bench Pro 59.2 against 58.4 and 64.0. A developer write-up of the platform documentation reports 750 billion total parameters with 31 billion active, a 512,000 token context window, a January 2026 cutoff, the model name k-exaone_v2.5, and no published price. LG is reviewing whether to apply NVIDIA's Dynamo inference software. Not established: rate limits, how long free lasts, independent evaluation of any score, and any open-weight plan.

How it got here

Aug 2026
LG AI Research releases the K-EXAONE model that 2.5 is measured against, in the second phase of the national foundation model project. The 70% improvement claim is relative to it.
Sep 22
Xiaomi releases the highest scoring open weights model in the world. Edition 031. Open weights, not a free interface, were the route rivals took.
Oct 25
A Chinese recruitment firm's lab publishes a 33 billion parameter coding model free. Edition 062. The second open-weight card on an Asian lab in 5 weeks.
Nov 9
At NVIDIA AI Day Seoul, LG AI Research opens the K-EXAONE 2.5 interface free to all comers, and names January as the date for a frontier-class model.

Who has to react

The Ministry of Science and ICT, which funds the programme and now has an adoption number to report; Korean developers choosing between this and foreign models; DeepSeek and Zhipu, named in LG's own comparison table; OpenAI, Anthropic and Google, whose Korean paid business this undercuts; and NVIDIA, whose conference hosted it and whose inference software LG is weighing.

What to watch

Would strengthen itThe Ministry of Science and ICT or LG publishes a developer, key or token adoption figure for K-EXAONE, or free access is still unmetered, by June 30, 2027.
Would weaken itFree access is withdrawn, priced or hard-capped below 1 million tokens a day, by June 30, 2027.
Would strengthen itLG AI Research ships K-EXAONE 3.0, or an EXAONE model it calls frontier-class, by March 31, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityEasy

PersistenceMonth

Cost is medium. Serving a 750 billion parameter model free is real compute spending every day, and forgone licence revenue on top, but the model was already trained and paid for. Commitment is medium: an open interface announced on a stage with a minister present is public, but it is a terms-of-service change, not a contract. Reversibility is easy and that is the honest score. A free tier can be metered, capped or priced in a single changelog entry, and LG has published no end date to hold it to. Persistence is months, because 2.5 is explicitly a verification model and LG has promised its successor in January.

The words

Foundation model: A large model trained once on broad data and then adapted to many tasks. Building one is a national-scale expense, which is why governments fund them.
Post-training: Extra training after the main run, to improve reasoning, instruction following and tool use. It is far cheaper than training from scratch.
Mixture of experts: A design where only part of the model runs for each request. It cuts the cost of serving a very large model.
Sovereign AI: A state's push to own models, chips and data centres inside its borders, so its language and industry are not dependent on foreign vendors.

Sources

Your call

Verify at LG's own release by Nov 20. Which explanation?

A call is final. Other people's calls show after yours.