Illustrative image: COEX in Seoul, where LG announced the free K-EXAONE 2.5 API at NVIDIA AI Day Seoul 2026. Image: Wikimedia Commons / Dr. Thomas Liptak, CC BY-SA 4.0.
📌 KEY HIGHLIGHTS
- Free from today: LG AI Research announced on 9 November 2026, at NVIDIA AI Day Seoul 2026, that the K-EXAONE 2.5 API is free — developers, startups, companies and research institutions can get an API key and use it for service development and research.
- The model: 750 billion parameters in total (31 billion active per token), text-only, Korean and English, with a 512K-token context window and a January 2026 knowledge cutoff.
- LG's headline claim: more than 70% better across 24 evaluation metrics than the K-EXAONE 2.0 model released in August — LG's own internal analysis, not independently verified.
- Agentic benchmarks (LG's figures): K-EXAONE 2.5 beats DeepSeek-V4.1-Flash on three of four tests, but trails GLM-5.3 on three of four — competitive, not leading.
- The catch: this is hosted API access only, not open weights — and LG has not said how long the free access lasts or what limits will follow.
What happened?
LG AI Research announced the free API on Monday 9 November 2026 at NVIDIA AI Day Seoul 2026, held at COEX in Seoul. From today, anyone — a solo developer, a startup, a company or a research institution — can obtain an API key through LG's own EXAONE API Platform and use K-EXAONE 2.5 for service development and research at no cost. LG says it opened the API to let people try a Korean-built AI model directly and apply it to their own services, and to gather real-world usage data before its next model.
K-EXAONE 2.5 is what LG calls a technology-validation model: it applies additional training to K-EXAONE 2.0, the second-generation model LG released in August 2026 as part of South Korea's Sovereign AI Foundation Model project. The point of this release is to prove that the training techniques — data design, post-training and reinforcement learning — actually work in practice, before LG builds them into its third-generation model. Kim I-reun, head of LG's EXAONE Lab, said internal analysis across 24 evaluation metrics showed performance up by more than 70% in just eight weeks of additional training. LG's co-director Lim Woo-hyung said the lab is now focused on completing a frontier-level EXAONE model by January 2027.
K-EXAONE 2.5 free API access: what you get
The model is a mixture-of-experts design: 750 billion parameters in total, with only about 31 billion active for each token, which is how LG keeps the cost of running it manageable. It handles text only, in Korean and English, and its context window holds up to 512K tokens — the combined input and output of a request cannot exceed 524,288 tokens. The full specification:
| Specification | K-EXAONE 2.5 |
|---|---|
| Parameters | 750 billion total, 31 billion active per token (mixture-of-experts) |
| Languages | Korean and English |
| Input | Text only — no image input |
| Context window | 512K tokens (524,288 combined input + output cap) |
| Knowledge cutoff | January 2026 |
| Access | Free API key via the EXAONE API Platform (docs.exaone.ai) |
| Endpoint | OpenAI-compatible POST /v1/chat/completions, model id k-exaone_v2.5 |
Because the endpoint follows OpenAI's format, existing code and agent tools can point at it by changing the base URL and model name — no new SDK to learn. The API supports chat, tool calling, structured outputs, code generation and streaming, and a reasoning-effort control (low, high or max, defaulting to max) sets how long the model thinks before answering. LG also ships agent guides for the Hermes general agent and the OpenCode coding agent. One detail matters for anyone building agents: the model returns tool-call decisions, but your own agent software executes the calls and manages the loop — the API does not run actions itself. The platform also lists EXAONE Tabular, a smaller model for predictions on structured data.
.jpg?width=900)
Illustrative image: server racks in a data centre — running a 750-billion-parameter model takes this kind of hardware, which is why LG is offering hosted API access rather than downloadable weights. Image: Wikimedia Commons / Carl Lender, CC BY 2.0.
What LG's benchmark numbers actually say
LG measured K-EXAONE 2.5 on open datasets and published four scores that test agentic AI — models that plan, use tools and complete real tasks on their own. The 24 metrics behind the 70% claim span nine areas, including long-context understanding, Korean long-form comprehension, instruction following and agentic tool use. Here are the four published agentic scores, with LG's own figures for the previous model and two Chinese rivals:
| Benchmark (agentic) | K-EXAONE 2.5 | K-EXAONE 2.0 | DeepSeek-V4.1-Flash | GLM-5.3 |
|---|---|---|---|---|
| SWE-Bench Pro (agentic coding) | 59.21 | 47.1 | 58.41 | 64.03 |
| BrowseComp (agentic search) | 78.0 | 50.87 | 75.8 | 79.0 |
| AutomationBench (SaaS tool use) | 74.6 | 16.9 | 74.9 | 71.6 |
| Claw-Eval (workplace problem-solving) | 77.29 | 62.11 | 75.78 | 77.64 |
The honest reading is two-sided. The jump over K-EXAONE 2.0 is large — AutomationBench leapt from 16.9 to 74.6 — so the eight weeks of extra training clearly did something real. Against the rivals, though, it is a draw: K-EXAONE 2.5 beats DeepSeek-V4.1-Flash on three of the four tests, but beats GLM-5.3 on only one. And every number above is LG's own measurement on public datasets; no independent evaluation has been published yet. That does not make the numbers wrong — it means they are a claim to test, not a verdict. Unlike Mistral's open-weight Large 4 release last month, where anyone can download the model and check it, K-EXAONE 2.5 can only be tested through LG's own API — which is exactly why the free access matters.
Why it matters for Indian developers
For Indian developers, the practical value is simple: a free, hosted 750-billion-parameter model with an OpenAI-compatible endpoint costs nothing to try and nothing to rewire for. If you already call OpenAI-style APIs, you can point the same code at LG's endpoint and compare it on your own tasks — code generation, debugging, refactoring and test-code writing are the uses LG names first, and agentic coding is where its benchmarks are strongest. English is fully supported, so the Korean-first origin is no barrier for Indian teams, though there is no Hindi support and no image input.
The bigger picture is the sovereign-AI race. South Korea is building its own frontier models with state backing, and LG's free API is how it gathers real-world usage before the third-generation model lands in January 2027. It is the same infrastructure logic driving the rest of the industry — the week Anthropic committed $3.4 billion to its own power supply for Claude, LG is spending access instead of money to win developers. If you want to understand what these models are actually being asked to do, read our explainer on what agentic AI actually means — K-EXAONE 2.5 is built for exactly that kind of work.
Should you care?
✅ If you build AI apps or coding agents: yes — a free 750-billion-parameter model with an OpenAI-compatible endpoint is worth an afternoon of testing, especially for agentic coding tasks where LG's numbers are strongest.
⏳ If you want a production dependency: wait — the benchmarks are LG's own, the free terms have no stated end date, and the model is text-only. Test it on your own workload; do not rebuild your stack around it yet.
FAQs
Is K-EXAONE 2.5 free to use?
Yes, for now. LG AI Research opened the K-EXAONE 2.5 API free of charge on 9 November 2026 through its EXAONE API Platform. Anyone can obtain an API key and use it for service development and research — but LG has not said how long the free access will last or what limits will apply afterwards.
What is K-EXAONE 2.5?
It is LG AI Research's 750-billion-parameter AI model, built as a technology-validation step for South Korea's Sovereign AI Foundation Model project. It applies additional training to K-EXAONE 2.0 (released August 2026) and is designed for agentic work such as coding, web search and automating tasks with software tools.
Is K-EXAONE 2.5 open source?
No. This release is hosted API access only — the weights are not downloadable. Earlier K-EXAONE models were released as open weights on Hugging Face under the Apache 2.0 licence, but LG has not announced open weights for version 2.5.
How do I get a K-EXAONE 2.5 API key?
Sign up on LG's EXAONE API Platform at docs.exaone.ai, obtain an API key, and call the OpenAI-compatible chat completions endpoint with the model name k-exaone_v2.5. A single request is capped at 524,288 tokens of combined input and output.
Takeaway
LG has put a 750-billion-parameter model in every developer's hands for free — not as charity, but as a live test of its sovereign-AI training before the next model arrives in January. The 70% gain and the benchmark wins are LG's own claims, and the free window has no promised end date, so treat this as a free trial worth running, not a permanent free tier. The real verdict will come from developers who test it on their own code.
If you build with AI tools, grab a free API key from LG's EXAONE API Platform and run one of your own coding tasks through it — your own result is worth more than any vendor benchmark.
Sources
- LG AI Research — EXAONE API Platform (model documentation and endpoint reference)
- The Korea Herald (Business) — LG AI Research opens K-Exaone 2.5 API for free (9 November 2026)
- Chosun Biz — LG AI Research opens K-EXAONE 2.5 API free, sees 70% performance gain (9 November 2026)
- Unite.AI — LG AI Research Opens Free API Access to Its K-EXAONE 2.5 Model (9 November 2026)
- Spectrum Web — LG opens K-EXAONE 2.5 API free for developers (9 November 2026)
Comments
Post a Comment