LG opens free K-EXAONE 2.5 API with noncommercial limits

LG AI Research opened free hosted access to K-EXAONE 2.5 on November 9, 2026. Developers can experiment through the EXAONE API Platform, but the current service terms restrict use to noncommercial purposes and prohibit integrations that let third parties call the model or submit inputs.

That restriction matters before building around the launch. A free API lowers the cost of evaluating a model. It does not, by itself, authorize a customer-facing application or commercial use of the resulting output.

What K-EXAONE 2.5 makes available

LG describes 2.5 as a technology validation model developed through additional training of the version released in August. The new announcement concerns access to a hosted service. It does not announce downloadable 2.5 weights.

The API documentation specifies a text-only model, an OpenAI-compatible chat completions endpoint and the model identifier k-exaone_v2.5. Its maximum context is 524,288 tokens, shared between the input and the requested generation budget. Exceeding that combined allowance returns an error.

Reasoning is always enabled. The reasoning_effort setting offers low, high and max levels, with max as the default. LG recommends low when speed matters, high for difficult math or coding, and max for more complex problems. The generation budget includes reasoning tokens, so it should not be treated as a promise of equally many visible answer tokens.

Read the noncommercial terms before integration

Section 2.1 of the platform terms forbids commercial use of either the model or its output. It also bars enabling third parties to call the model or submit inputs through products or services. The terms state that a separately released model’s weight license does not replace the platform’s own rules.

Users must be at least 19. API keys expire 30 days after issuance, and LG reserves the ability to change or discontinue models. The terms also allow future charges after fees are posted. Plan an evaluation around the current conditions rather than assuming permanent free access or a production service commitment.

Free access still has request limits

The rate limits documentation explains how response headers report request allowances and remaining usage over second, minute and day windows. Exceeding a limit returns HTTP 429. Its numerical header example illustrates the format. It is not a universal quota guarantee.

For an experiment, record actual response headers and handle HTTP 429 responses with an appropriate delay rather than immediate repeated retries. Check the current allowance before estimating how long a batch evaluation will take.

Check consent before sending real data

LG’s November 9 privacy policy describes input and output retention for model training and evaluation as optional processing based on consent. Without that consent, it says the content is processed to handle the request and is not stored afterward.

The policy distinguishes that content from required usage metadata, which can remain for a year after account deletion. Consent can be withdrawn in My profile under Consent and Notification. LG also says material already incorporated into model training cannot technically be separated out and deleted.

Start with synthetic or public test material. A content-retention choice is only one part of deciding what can be submitted, and the service terms separately prohibit including another person’s personal information in an input.

LG’s agent scores are vendor results

LG reports 74.6 on AutomationBench-Openset, 59.2 on SWE-Bench Pro, 78.0 on BrowseComp and 77.3 on Claw-Eval. These are its own measurements on public datasets, covering automation, software work, web research and workplace agent tasks. The announcement compares the model with DeepSeek-V4.1-Flash and GLM-5.3.

ByteForward has not independently reproduced those scores. A useful evaluation would measure completion accuracy on your own permitted tasks, total tokens, response time and failure cases together. Keep the reasoning setting fixed when comparing runs, then vary it deliberately to see whether extra reasoning improves the outcome enough to justify the delay.

Hosted evaluation and self-hosting are separate choices

The hosted 2.5 API is a different deployment path from the NVIDIA Dynamo recipes for Korean models covered earlier. That package includes a recipe for K-EXAONE 2.0 and targets teams operating their own GPU infrastructure. Its hardware configuration and licensing context should not be carried over to this hosted service.

For now, the practical opportunity is a bounded evaluation: confirm the intended use is permitted, choose an appropriate dataset, record the model and settings, and inspect the outputs. A positive test result still leaves commercial permission and deployment reliability to resolve before considering any commercial deployment.

Featured image: ByteForward editorial illustration. LG AI Research logo: LG AI Research. No endorsement implied.