LG K-EXAONE 2.5 Is Free: 750B Korean AI Model You Can Call Today

LG AI Research put its biggest model in developers' hands for free on November 9, 2026. K-EXAONE 2.5 is a 750-billion-parameter model built for Korean and English, and anyone can now call it through the EXAONE API Platform with an API key. No subscription, no per-token bill, no credit card step in the way. That makes it one of the cheapest ways to test a frontier-class model right now, because the price is zero. The catch is the part LG has not answered: nobody knows how long free lasts, and the model has no published open weights and no independent benchmark results yet.

This article covers what LG actually announced, what the model can and cannot do, what its own benchmark numbers mean, and how to try it in about five minutes.

What LG actually announced

LG AI Research announced the move at NVIDIA AI Day Seoul 2026, held at COEX in Seoul on November 9, 2026. The announcement opens the EXAONE API Platform to outside developers for the first time, with K-EXAONE 2.5 as the flagship model on it. Developers, startups, companies, and research institutions can obtain an API key and use the model for service development and research at no cost, according to LG's announcement as reported by the Asia Business Daily and Chosun Biz.

The platform lists two models today: K-EXAONE 2.5, described by LG as its flagship agentic model, and EXAONE Tabular, a smaller model for predictions on structured data. The platform's own pages confirm both endpoints exist, along with documentation for rate limits and a changelog.

K-EXAONE 2.5 is a technology-validation model. LG built it by applying additional training to the second-generation K-EXAONE model it released in August 2026 under South Korea's Sovereign AI Foundation Model project. LG says the training techniques validated here will feed into a third-generation model, and it aims to complete a frontier-level EXAONE model by January 2027.

What K-EXAONE 2.5 is

The model is a mixture-of-experts design with 750 billion total parameters and 31 billion active per token. It handles text only, in Korean and English, with a 512K-token context window and a knowledge cutoff of January 2026. Its documentation lists chat, tool calling, structured outputs, code generation, and streaming as supported features.

Two details in that table deserve attention. First, the active parameter count went down from 37 billion to 31 billion while total parameters rose, which fits a model that routes each token through fewer experts. Second, the earlier K-EXAONE 2.0 shipped as open weights under Apache 2.0, while 2.5 is API-only so far. LG has not announced open weights for 2.5, so anyone who wants to self-host is still waiting.

The benchmarks LG is quoting

LG published results on four agentic benchmarks, measured internally on open datasets. The company says the figures show K-EXAONE 2.5 competing with DeepSeek V4.1 Flash and GLM-5.3, two models that already have public pricing. Here is the table from LG's own documentation:

Read the table carefully, because the headline version is flattering. K-EXAONE 2.5 beats its predecessor on all four benchmarks, and the jump on AutomationBench, from 16.9 to 74.6, is the largest single gain. Against the two rivals, the picture is mixed: it edges DeepSeek V4.1 Flash on three of the four benchmarks and trails it slightly on AutomationBench, while GLM-5.3 leads it on SWE-Bench Pro and BrowseComp and trails it only on AutomationBench. LG's claim that the model is competitive with both is fair on these numbers, but "competitive" is not the same as leading, and every number in the table comes from LG's own testing. No independent lab has published results for K-EXAONE 2.5 yet.

LG also says overall performance across 24 evaluation metrics improved by more than 70 percent compared with the August model, in its own internal analysis. That figure comes from Kim I-reun, head of the EXAONE Lab at LG AI Research. A 70 percent gain across 24 metrics in eight weeks is a large claim, and it is LG's own measurement, so treat it as the company's report rather than a verified result.

What free actually means

The free part is simple: the EXAONE API Platform currently charges nothing for K-EXAONE 2.5, and LG has not published a price list for it. The platform's changelog describes the service as currently free of charge. What LG has not said is how long that lasts, what rate limits apply, or what the paid price will be when free ends. That is the main pricing risk. Anyone building a service on the free API should assume a price will appear and plan for it, because LG has given no commitment that free access is permanent.

For comparison, here is what the two models in LG's own benchmark table cost on their public APIs:

Rates are approximate conversions at November 9, 2026 market rates (1 USD = 0.748 GBP, 1 USD = 1.412 CAD). DeepSeek's rates double during peak hours, and its API bills from 00:00 Beijing time on November 1, 2026 for the new Search and Fetch features. GLM-5.3's rates are from Z.ai's official pricing page, corroborated by independent pricing trackers. For a closer look at DeepSeek's pricing, see our DeepSeek V4.1 Flash 1017 review.

The comparison matters because it shows what free is worth. A million input tokens on GLM-5.3 costs $1.40, and a million output tokens costs $4.40. The same work on K-EXAONE 2.5 costs nothing today. That gap is real money for anyone running agents that chew through tokens, but it only holds while LG keeps the free tier open.

How to try it

Getting started takes an API key and an OpenAI-compatible client. The endpoint follows the standard chat completions format, so existing code needs little more than a new base URL and model name.

1. Open the EXAONE API Platform documentation and create an account.
2. Generate an API key from the platform.
3. Point your client at the platform's chat completions endpoint and set the model to k-exaone_v2.5.
4. Send a request. Input tokens plus the max_tokens allowance, which includes reasoning, cannot exceed 524,288 tokens; larger requests return a 400 error.

Two settings are worth knowing before the first call. The reasoning_effort parameter defaults to max, which gives the model its longest thinking time and highest token use. Lower it for short tasks. The preserve_thinking option can carry earlier reasoning into multi-turn requests, which helps agents that build on prior steps. LG's agent guide includes configurations for the Hermes general agent and the OpenCode coding agent, and the model returns reasoning and tool-call decisions through the API while the surrounding agent harness executes the actions.

What it cannot do yet

The limits are as important as the specs. K-EXAONE 2.5 is text-only, so it cannot see images, hear audio, or generate pictures. It works in Korean and English only, which narrows its use for anyone working in a third language. There are no published open weights for 2.5, so self-hosting and fine-tuning are not options today; the earlier 2.0 release remains the open-weight choice. LG has published no rate limits in plain numbers on its public pages, and no paid price, so heavy users have no way to budget for the day free access ends. Finally, every benchmark number so far comes from LG itself. Until an independent evaluation appears, the honest position is that the model looks competitive on LG's own tests and unproven everywhere else.

The Korean-language focus is also a real trade-off in the other direction. LG describes K-EXAONE 2.5 as a model that understands Korean users well, and Korean is where it should have an edge over models trained mostly on English data. For Korean-language agents, customer support, and coding tools, that is the most interesting part of this release. For English-only work, the model is one more capable option in a crowded field, and its main advantage is the price.

Who this is actually for

Three groups get the most from free access. Developers building Korean-language agents can test a large model without a budget approval. Teams comparing agentic coding models can run K-EXAONE 2.5 against DeepSeek V4.1 Flash and GLM-5.3 on their own tasks, using the same OpenAI-compatible tooling for all three. Researchers can probe a sovereign-AI model from a major Korean lab at no cost. If none of those apply, the release is still worth a look, because a 750-billion-parameter model with a 512K context window and zero cost is rare, and the free window may not last.

Common mistakes to avoid

Do not build a production service on the free API without a fallback. LG has said nothing about an end date, a price, or a service level, so a paid migration could arrive with little warning. Do not assume the open-weights license from K-EXAONE 2.0 applies to 2.5. It does not; 2.5 is API-only, and the terms of service on the platform govern its use. Do not leave reasoning_effort at its default of max for every call. It is the right setting for hard reasoning tasks, but it burns tokens and time on simple ones. And do not quote the benchmark table as proof the model beats its rivals. It beats DeepSeek on three of four benchmarks and trails GLM-5.3 on two, and all four numbers are LG's own.

Is K-EXAONE 2.5 really free?

Yes, for now. LG opened the API on November 9, 2026 with no charge, and the platform's changelog describes the service as currently free. LG has not said when free access ends or what it will cost after that.

Can I download the model weights?

No. LG has published weights for earlier K-EXAONE models, including K-EXAONE 2.0 under Apache 2.0, but it has not announced open weights for version 2.5. Access is through the API only.

What languages does it support?

Korean and English, text only. The model does not accept images or audio, and it does not generate images.

How does it compare with DeepSeek V4.1 Flash?

On LG's own benchmarks, K-EXAONE 2.5 scores slightly higher than DeepSeek V4.1 Flash on SWE-Bench Pro, BrowseComp, and Claw-Eval, and slightly lower on AutomationBench. DeepSeek costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, while K-EXAONE 2.5 is free for now. DeepSeek's numbers are also company-reported, so neither side has independent confirmation.

How do I call the API?

Get a key from the EXAONE API Platform, then send a standard chat completions request with the model name k-exaone_v2.5. The endpoint is OpenAI-compatible, so most existing clients work with a base URL change. Keep total input plus generation under 524,288 tokens.

Should I switch from GLM-5.3?

Not on price alone. GLM-5.3 has a 1M-token context window, published open weights, and a known price of $1.40 per million input tokens and $4.40 per million output tokens. K-EXAONE 2.5 is free and has a 512K context window, but no weights and no price commitment. Test both on your own workload before deciding.

LG's free opening is a genuine chance to test a large Korean model without spending anything, and the benchmark numbers give it a fair claim to sit at the same table as DeepSeek V4.1 Flash and GLM-5.3. The honest version is shorter: the model looks strong on its maker's own tests, the price is zero with no promised end date, and the proof that matters will come from independent testing and real use. Try it while it is free, keep a fallback ready, and check the platform's terms before building anything you depend on.

Sources

• EXAONE API Platform — LG AI Research, official documentation (platform, models, rate limits, changelog; verified loading November 9, 2026)
• LG AI Research — official site (verified loading November 9, 2026)
• LG AI Research Opens Free API Access to Its K-EXAONE 2.5 Model — Unite.AI, November 9, 2026 (benchmark table from LG documentation, endpoint reference, model specs)
• LG Unveils K-EXAONE 2.5, Boosts Agentic AI Performance by 70% — Asia Business Daily, November 9, 2026 (announcement at NVIDIA AI Day Seoul 2026, free API, 70% improvement claim)
• LG AI Research opens K-EXAONE 2.5 API free, sees 70% performance gain — Chosun Biz, November 9, 2026 (70% figure from Kim I-reun, EXAONE Lab head; AutomationBench 74.6 vs GLM 71.6)
• LG opens K-EXAONE 2.5 API free for developers — Spectrum Web, November 9, 2026 (platform changelog "currently free of charge", no end date stated, no open weights for 2.5, January 2027 frontier goal)
• LG Opens DeepSeek-Level Agentic AI K-EXAONE Free to Public — NoCut News, November 9, 2026 (API key access for developers, startups, companies, research institutions)
• GLM-5.3 benchmarks and pricing — The Model Gap (GLM-5.3 at $1.40 input / $4.40 output per 1M tokens, official pricing source)
• DeepSeek V4.1 Flash 1017: The $0.15 Coding Model and Its New Search API — AI Core Updates, October 23, 2026 (DeepSeek off-peak rates, peak doubling, Search/Fetch billing)
• Kimi K3.1 Pro Is Live: $2 API, 4th on AI Index, Open Weights Nov 4 — AI Core Updates, October 24, 2026 (open-weight model comparison context)
• Tencent Hy4 Is Live: 770B Model That Sees and Uses Your PC — AI Core Updates, November 8, 2026 (large-model launch context)