LG just made its 750 billion parameter model free to try, and the weights are still not yours
Hello,
LG made its big model free to try today. Not free to own. Free to try.
The model is K-EXAONE 2.5, and LG AI Research opened it through the EXAONE API Platform on November 9, at NVIDIA AI Day in Seoul. You sign up, you get an API key, and you can call the model. The platform says the service is free right now. LG has not said how long that lasts, or what the limits will be when it ends.
It is a big model. 750 billion parameters in total, with 31 billion active for each token. It reads and writes in Korean and English only, and it works with text only. The context window is 512K tokens, so it can hold a lot of text in one go. It does chat, tool calling, structured outputs, code work and streaming, and the endpoint fits the OpenAI shape, so existing code can point at it with the name k-exaone_v2.5.
Here is what free buys you, and what it does not.
Free buys you a real test. You can run your own coding jobs, your own searches, your own agent work, and see how it holds up outside a demo.
Free does not buy you the model. This release is API only. LG put weights out for earlier EXAONE models, but it has not said it will do that for 2.5. You can use it. You cannot download it, fine-tune your own copy, or run it on your own machines.
Then the scores. LG says 2.5 is up more than 70 percent on the August model across 24 tests, in about eight weeks of extra training. On its own numbers it scores 59.21 on SWE-Bench Pro, 78 on BrowseComp, 74.6 on AutomationBench and 77.29 on Claw-Eval, close to DeepSeek V4.1 Flash and GLM 5.3. Those are LG's numbers, from LG's own tests. Nobody outside LG has checked them yet.
LG says the point of the free API is to learn from real use before the next model, a third-generation EXAONE it wants at frontier level by January 2027. That makes sense. A free API is a cheap way to find out what breaks in real work.
So this week gave us two kinds of open. Yesterday, a model you can download but can hardly run. Today, a model you can run for free but cannot keep. If you try the LG one, tell me what it does well with Korean, or with code. That is the part I am most curious about.
Cover image: AI-generated