Key takeaways

  • LG AI Research's API docs, with a changelog entry dated November 9, 2026, list K-EXAONE 2.5 and say the service is currently free.
  • The model ID is k-exaone_v2.5 on an OpenAI-compatible chat completions endpoint, with a 512K context and 750B parameters (31B active).
  • No K-EXAONE 2.5 weights are on Hugging Face as of November 9; the K-EXAONE 2.0 models are.

LG AI Research opened the K-EXAONE API to developers on November 9, and its API documentation says the service is free for now. The model is K-EXAONE 2.5, a 750-billion-parameter text model you call with an API key and an OpenAI-style chat completions request. There are no downloadable weights for it yet.

This page covers what the K-EXAONE API is, how to call it, what the docs say about limits and cost, and how it relates to the K-EXAONE models on Hugging Face.

What the K-EXAONE API is

The API is part of LG AI Research's EXAONE API Platform at docs.exaone.ai. Its changelog has one entry, dated November 9, 2026, saying the EXAONE models are available through the API. It lists K-EXAONE 2.5 as the language model. The platform also lists a second model, EXAONE Tabular, for predictions on structured data.

The docs describe K-EXAONE 2.5 as a flagship agentic model that understands Korean users well. "Agentic" here means built for tasks where the model calls tools and works through several steps.

Is the K-EXAONE API free?

Yes, for now. The changelog says, in Korean, that the service "is currently provided free of charge." The Korea Times reported the same on November 9, and so did other Korean outlets. The docs give no end date for free use and list no paid tiers or per-token prices. If LG sets a price, it has not published one as of November 9.

The rate limits page explains how limits work: they are counted per second, per minute and per day, and going over returns an HTTP 429 error. It gives no limits for an account. The numbers in its example are only samples of the header format. Your real limits show in the x-minute-rate-limit and x-day-rate-limit response headers.

How to call the K-EXAONE API

You get an API key from the docs site, using the "issue API key" button. The model ID is k-exaone_v2.5. Requests go to:

POST https://api.exaone.ai/v1/chat/completions

The API is OpenAI compatible. Authentication is a bearer token in the Authorization header, using your EXAONE_API_KEY. The docs show examples in cURL, Python and JavaScript, with streaming and without.

A few settings matter:

  • reasoning_effort takes low, high or max, and defaults to max. The docs suggest low for fast replies and high for math and coding.
  • Reasoning is always on. The model's reasoning comes back in reasoning_content and the answer in content. Reasoning tokens are counted in usage.reasoning_tokens and count toward max_tokens.
  • For agents and multi-turn chats, the docs recommend chat_template_kwargs.preserve_thinking=true so earlier reasoning is kept.
  • The recommended defaults are temperature 1.0 and top_p 0.95.

Model specs from the docs

Item Value
Model ID k-exaone_v2.5
Parameters 750B total, 31B active
Context length 524,288 tokens (listed as 512K)
Modality Text only
Languages Korean and English
Features Chat, tool calling, structured outputs, code generation, streaming
Knowledge cutoff January 2026

If input tokens plus max_tokens go over the context length, the API returns a 400 error.

How K-EXAONE 2.5 scores

LG's docs compare K-EXAONE 2.5 with K-EXAONE 2.0, DeepSeek-V4.1-Flash and GLM-5.3. These are LG's own numbers, not independent tests.

Benchmark 2.5 2.0 DeepSeek-V4.1-Flash GLM-5.3
SWE-Bench Pro (coding) 59.21 47.1 58.41 64.03
BrowseComp (search) 78 50.87 75.8 79
AutomationBench (tool use) 74.6 16.9 74.9 71.6
Claw-Eval (tool use) 77.29 62.11 75.78 77.64

The jump over 2.0 is large, most of all on AutomationBench. Against the two rivals it is mixed: ahead of DeepSeek on three of the four, behind GLM-5.3 on all four. The Korea Times reports that LG says performance across 24 benchmarks rose more than 70 percent over the earlier model. The article does not list those 24.

K-EXAONE on Hugging Face

People searching for "K-EXAONE Hugging Face" are looking for weights, and 2.5 has none yet. When we checked on November 9, the LGAI-EXAONE page listed K-EXAONE 2.0 (750B-A37B, plus FP8, NVFP4 and DSpark variants) and the earlier K-EXAONE-236B-A23B. A search of Hugging Face for K-EXAONE-2.5 returned no models. LG has not said whether it will release them. For now, 2.5 is hosted only.

What to watch

The Korea Times says LG plans to apply the training methods tested in 2.5 to its next model, and that its co-head Lim Woo-hyung said the lab is aiming for a frontier-level EXAONE model by January. Watch the changelog for a price or an end to free access, and the Hugging Face page for any 2.5 weights.

Frequently asked questions

Is the K-EXAONE API free?

LG's changelog says the service is currently free. It gives no end date and lists no prices, so that may change.

What is the K-EXAONE API model name?

The model ID is k-exaone_v2.5. Send it to https://api.exaone.ai/v1/chat/completions with a bearer token.

Is K-EXAONE on Hugging Face?

K-EXAONE 2.0 and the 236B model are on the LGAI-EXAONE Hugging Face page. K-EXAONE 2.5 weights were not listed as of November 9, 2026.

What is the difference between K-EXAONE 2.0 and 2.5?

LG's docs show 2.5 scoring higher on all four benchmarks listed, for example 74.6 versus 16.9 on AutomationBench. These are LG's own results.

Sources

  1. K-EXAONE 2.5 API documentationLG AI Research
  2. EXAONE API Platform changelogLG AI Research
  3. EXAONE API Platform modelsLG AI Research
  4. Rate limitsLG AI Research
  5. LGAI-EXAONE on Hugging FaceHugging Face
  6. LG AI Research opens K-EXAONE 2.5 API free of chargeThe Korea Times