LG AI Research has opened free API access to K-EXAONE 2.5, its 750-billion-parameter flagship model, putting Korea’s sovereign AI in the hands of any developer who wants it. The lab announced the move on 9 November at NVIDIA AI Day Seoul, launching the EXAONE API Platform with no charge for now. The model handles Korean and English text, carries a 512,000-token context window, and on LG’s own measurements now matches DeepSeek V4.1 Flash and GLM-5.3 on agentic benchmarks. LG says the model improved more than 70 percent in eight weeks, and it is aiming for a frontier-level EXAONE model by January 2027.
## What LG actually opened up
The EXAONE API Platform offers two models through one hosted service: K-EXAONE 2.5, LG’s flagship agentic model, and EXAONE Tabular, a smaller model for predictions on structured data. Developers, startups, companies and research institutions can obtain an API key and use the service for development and research without paying.
LG has not said how long the free period will last, or what limits will apply once it ends. That silence matters: a free tier that turns paid without warning changes the maths for any team building on it, so developers will want the terms published before they commit real workloads.
The endpoint is OpenAI-compatible, so existing client code can point at it with the model name k-exaone_v2.5. It supports chat, tool calling, structured outputs, code generation and streaming, and exposes a reasoning_effort control set to low, high or max by default.
Input tokens plus the generation allowance cannot exceed 524,288 tokens. The model is text-only, works in Korean and English, and has a January 2026 knowledge cutoff — a narrower scope than the multimodal flagships it is measured against, but a fair one for a model built first for Korean users.
## The benchmark claims, in context
LG compares K-EXAONE 2.5 with its predecessor and two outside systems on four agentic benchmarks, all measured by LG itself on open datasets. It scores 59.21 on SWE-Bench Pro, ahead of DeepSeek V4.1 Flash at 58.41 but behind GLM-5.3 at 64.03.
It scores 78.0 on BrowseComp, 74.6 on AutomationBench — up from 16.9 for K-EXAONE 2.0 — and 77.29 on Claw-Eval. These are self-reported figures, not independent results, and the AutomationBench leap is the one that most invites independent outside checking.
Kim I-reun, head of LG’s EXAONE Lab, said internal analysis across 24 evaluation metrics showed overall performance up more than 70 percent on the August model, achieved in about eight weeks. That pace is the claim LG wants remembered, and it is also the claim most exposed if outside tests disagree.
[DeepSeek’s free models](https://geroundtable.com/deepseek-nearly-matches-daybreak-cyber/) have set a high bar for open-weight rivals, and [GLM-5.3’s own cyberattack-test results](https://geroundtable.com/glm-5-3-cyberattack-tests-open-weight/) drew attention for the opposite reason. K-EXAONE 2.5 now has to prove it belongs in that company, on other people’s tests rather than its own.
## Why It Matters
Free access to a 750-billion-parameter model with a half-million-token context is a genuine opening for developers outside Korea, who have had little chance to test a Korean frontier contender. Sovereign AI has mostly been a government story until now; LG is turning it into a developer story by letting anyone build on the model.
That matters because the next EXAONE is supposed to be frontier-level, and real usage is the fastest way to find its weaknesses. A model that only its makers have run is a press release; a model thousands of developers have pushed is evidence.
The opening has limits worth noting. This release is API access only: LG has published weights for earlier K-EXAONE models, but it has announced no open weights for version 2.5. That is a different bet from [Mistral, which set its trillion-parameter Large 4 free as open weights](https://geroundtable.com/mistral-large-4-public-open-weight-release/).
LG is also considering NVIDIA’s Dynamo inference software to improve serving efficiency, though that remains a possible step rather than a confirmed deployment. The free tier, the API-only release and the Dynamo study all point the same way: LG wants usage first, and openness later.
## Our Read
LG’s play is distribution, not openness. A free API buys what benchmarks cannot: developers actually running the model on their own tasks, in Korean and English, and finding where it breaks. The 70 percent gain in eight weeks is a strong claim, but it is LG grading its own homework.
The honest test comes when outside teams publish their own numbers. If adoption stays thin, the free tier will look like a showcase; if it spreads, LG gains a feedback loop its rivals pay for. And a feedback loop is exactly what a team racing toward a January frontier deadline needs most.
The timing is also the story. LG announced at NVIDIA AI Day Seoul, in front of the hardware audience that decides which models get served at scale, and it set a public deadline: a frontier-level EXAONE by January 2027. That is eight weeks away.
[Thinking Machines’ Inkling 2 Small](https://geroundtable.com/thinking-machines-inkling-2-small/) shows the other route, a huge model whose weights are still locked behind safety testing. LG has chosen the opposite order, access first and openness later, and January will show whether the bet paid off.
## Background
K-EXAONE 2.5 is a technology-validation model built by further training the second-generation model LG released in August 2026, which shipped as an open-weight Apache 2.0 release with 37 billion active parameters and a 262,144-token context. The work sits inside Korea’s sovereign AI foundation model project, and LG says the training methods validated here will feed the third-generation model in development.
Source: [EXAONE API Platform](https://docs.exaone.ai/) and [MoneyToday report](https://www.mt.co.kr/en/industry/2026/11/09/2026110909413580792)
0 Comments