K-EXAONE 2.5 free API: LG's 750B model rivals DeepSeek and GLM on agentic benchmarks

LG AI Research has put its biggest model in developers' hands — for free. Announced November 9 at NVIDIA AI Day Seoul 2026, the K-EXAONE 2.5 API is now open to anyone who signs up for an API key on the EXAONE API Platform. No pricing tier, no waitlist: developers, startups, companies, and research institutions can use the 750-billion-parameter model for service development and research at no cost, starting today.

That is a notable shift for a lab that, until now, reached outside developers mainly through open weights. K-EXAONE 2.0, released in August 2026, shipped as a 750B-parameter model under Apache 2.0 — but version 2.5 is a hosted API, not a weights release. LG has not announced open weights for 2.5, and has not said how long the free access will last. What it has said is that the free period is a deliberate test: the lab wants to see how the model performs in real service environments before the third-generation model arrives.

A technology-validation model#

K-EXAONE 2.5 is what LG calls a technology-validation model: the August model with additional post-training applied, built to prove out the training techniques that will go into the third-phase EXAONE. The architecture is a sparse mixture-of-experts — 750 billion total parameters, with 31 billion active on any given token — which is why it can be served at a cost LG is willing to eat for now. It handles text only, in Korean and English, with a 512K-token context window and a January 2026 knowledge cutoff.

Using it should feel familiar. The endpoint is OpenAI-compatible — POST /v1/chat/completions with model ID k-exaone_v2.5 — with chat, tool calling, structured outputs, code generation, and streaming. The docs include agent guides for the Hermes general agent and the OpenCode coding agent, and expose reasoning controls directly: reasoning_effort defaults to max, and preserve_thinking can carry earlier reasoning into multi-turn requests. The main constraint is the context budget: input tokens plus the generation allowance cannot exceed 524,288 tokens per request.

The numbers — and the caveats#

LG's pitch is agentic performance. On its own measurements, run on open datasets, K-EXAONE 2.5 scores 74.6 on AutomationBench (SaaS tool-use automation), 59.2 on SWE-Bench Pro (agentic coding), 78.0 on BrowseComp (agentic web search), and 77.3 on Claw-Eval (agent problem-solving in work environments). Against the two Chinese frontier open-weight models LG chose as its comparison set, that is a near-dead heat: DeepSeek-V4.1-Flash scores 74.9, 58.4, 75.8, and 75.8 on the same four, while GLM-5.3 scores 71.6, 64.0, 79.0, and 77.6. LG also says the model improved more than 70% over the August version across a combined assessment of 24 evaluation metrics.

Every one of those numbers is LG's own. There is no independent evaluation yet, and the comparison set is LG's choice — two rivals, on four benchmarks the lab picked. The honest reading: this is a 750B model that, on LG's scoreboard, now belongs in the same conversation as DeepSeek and GLM for agent work, and the free API means anyone can test that claim without paying for the privilege. The benchmark table above is LG's chart, reproduced from its announcement.

What to watch#

Three things to watch. First, whether independent evaluations confirm the numbers — a 70% jump over the August model is the claim that will draw the most scrutiny, and the free API makes it checkable. Second, what the free period turns into: LG has not named a price, a rate limit, or an end date, and the platform also lists a smaller EXAONE Tabular model, suggesting a broader hosted lineup is coming. Third, the road to the third-phase model — LG says it is reviewing NVIDIA's open-source Dynamo inference technology to speed the model up, and aims to complete a frontier-class EXAONE by January 2027.

For now, the practical upshot is simple: a 750B-parameter model from a major Korean lab is one API key away, free, through an endpoint your existing OpenAI-compatible code already speaks. The weights are not open this time — but the trial is. And with open-weight giants like Kimi K3.1 Pro and DeepSeek's 284B model running on 64GB of RAM already competing for the same developers, LG's free API is less a giveaway than a bid to be tested.

Sources: LG AI Research, ‘K-EXAONE 2.5 API 무료 공개’ (November 9, 2026), here; DigitalToday, ‘LG AI Research opens K-EXAONE 2.5 API for free, says performance up 70 percent’ (November 9, 2026), here; Unite.AI, ‘LG AI Research Opens Free API Access to Its K-EXAONE 2.5 Model’ (November 9, 2026), here; EXAONE API Platform documentation, here. Cover and inline graphics are from LG AI Research.