K-EXAONE 2.5: LG opens free API with 70% boost

LG AI Research opened today the K-EXAONE 2.5 API for free. The model with 750 billion parameters (31 billion active) had performance 70% higher than the August version, in just eight weeks. The announcement was made at NVIDIA AI Day Seoul 2026, on November 9.

In this article you will see:

• What K-EXAONE 2.5 is and who can use it
• The official numbers on the four agentic AI benchmarks
• How the model compares to DeepSeek and GLM
• How to test K-EXAONE 2.5 today

What is K-EXAONE 2.5, LG's open model

K-EXAONE 2.5 is a language model from LG AI Research, with 750 billion total parameters and 31 billion active. It is a technical validation version, created with additional training on the model released in August 2026.

The model's focus is agentic AI: planning, tool use and task execution in work environments. LG says the model supports complex reasoning, code generation, debugging, refactoring and workflow automation.

The differentiator of this release is access: the API is available for free to developers, startups, companies and research institutions. Just get an API key on the EXAONE platform. There is no usage cost, nor an announced time limit.

The model is text only, in English and Korean, with knowledge cutoff in January 2026. It does not process images, audio or video.

The official numbers for K-EXAONE 2.5 on benchmarks

LG measured the model on four agentic AI benchmarks, using public datasets. The results are internal, not audited by third parties:

• AutomationBench-Openset: 74.6 (task automation with SaaS tools)
• SWE-Bench Pro: 59.2 (development and modification of real software)
• BrowseComp: 78.0 (web search and navigation)
• Claw-Eval: 77.3 (problem solving in work environment)

The company says aggregate performance across 24 evaluation metrics improved more than 70% compared to the August model. That number was measured internally, in eight weeks of work.

K-EXAONE 2.5 vs DeepSeek and GLM

On the four benchmarks, LG says K-EXAONE 2.5 was competitive with open frontier models, including DeepSeek-V4.1-Flash and GLM-5.3 (max). The comparison is done by LG itself, with its own tests.

There is no public table with competitor scores side by side. The competitiveness claim is from the company, not from an independent ranking. Anyone who wants to validate needs to run the model on their own infrastructure.

This is a common pattern in recent launches: the company publishes its own numbers and lets the community test. Anyone who evaluates agent benchmarks knows that internal results need independent reproduction before becoming fact.

If you want to go deeper into how to compare these numbers, it's worth reading the 7 tips for evaluating agentic AI benchmarks.

The partnership with NVIDIA and the plan for January

LG also announced technical cooperation with NVIDIA. The idea is to apply NVIDIA's Dynamo technology to improve inference efficiency and service scalability.

The company said it plans to apply what it learned from K-EXAONE 2.5 to develop a next generation model by January 2027. In other words, this release may be short lived. A technical validation model, not a final version.

This matters for anyone planning to use the model in production: a free model with no end date can change price or disappear. LG did not say how long free access will be available.

Limitations you need to know

K-EXAONE 2.5 has clear limitations. It is text only, with no support for images, audio or video. It works in English and Korean. The knowledge cutoff is January 2026, which can limit tasks that require recent information.

The benchmarks are measured by LG itself, in its own tests, with public datasets. There is no independent audit. Scores can vary on other workloads.

For those already testing open models, it is worth comparing with the Hy4 from Tencent, which also has free access and public numbers. If you follow CLI releases, you can see how track Codex CLI releases without breaking your workflow.

How to test K-EXAONE 2.5 today

1. Access the EXAONE platform at docs.exaone.ai.
2. Request a free API key.
3. Use the OpenAI API compatible endpoint to send requests.
4. Test on coding, automation and web search tasks.

The API is compatible with OpenAI's format, which makes integration with existing tools easier. Anyone already using OpenAI SDKs can swap the endpoint and test with little effort.