K-EXAONE 2.5 Free API Explained: Try LG’s 750B AI Model Without Paying
K-EXAONE 2.5 is now free to try. On November 9, 2026, LG AI Research opened its flagship model to anyone with an API key, through the new EXAONE API Platform — no payment, no waitlist, no credit card mentioned in the announcement. The model behind the offer is serious hardware-scale AI: 750 billion parameters in total, with 31 billion active on any given token, a 512,000-token context window, and agentic benchmark scores that LG says match DeepSeek V4.1 Flash and GLM-5.3. This guide explains what K-EXAONE 2.5 actually is, what the free API covers, how to call it from your own code, and the honest limits you should know before you build anything on it.

Table of Contents
- What K-EXAONE 2.5 Is
- The Free API: What LG Actually Opened Up
- K-EXAONE 2.5 Specs at a Glance
- How the K-EXAONE 2.5 Free API Works
- K-EXAONE 2.5 Benchmarks: What LG Claims — and What It Does Not Prove
- What You Can Build With K-EXAONE 2.5
- K-EXAONE 2.5 vs K-EXAONE 2.0: What Changed
- The Sovereign AI Story Behind K-EXAONE 2.5
- K-EXAONE 2.5 vs DeepSeek V4.1 Flash and Other Free Options
- Honest Limitations to Know Before You Commit
- How to Start Using the K-EXAONE 2.5 Free API in 10 Minutes
- Who K-EXAONE 2.5 Is Best For — and Who Should Skip It
- The Verdict
- FAQs About K-EXAONE 2.5
What K-EXAONE 2.5 Is
K-EXAONE 2.5 is LG AI Research’s newest large language model, and the second-generation flagship of South Korea’s sovereign AI programme. LG announced the free API on November 9, 2026, at NVIDIA AI Day Seoul 2026, held at the COEX convention centre in Seoul. From that day, developers, startups, companies, and research institutions can obtain an API key from the EXAONE API Platform and use the model in their services and research at no cost.
The “K” in the name matters. K-EXAONE is the Korean national foundation-model line, built under a government-backed sovereign AI project whose goal is to reduce the country’s reliance on foreign AI systems. The model is text-only and works in Korean and English. For a US developer, the Korean origin is not the point — the point is that a 750-billion-parameter model with a half-million-token context window is now callable for free, through an interface your existing code probably already speaks.
LG describes the model as a technology-validation model. That is deliberate language. It means this release exists to prove that the underlying model works in real services and to gather real-world usage data before the next step. LG applied additional training to the model it released in August 2026, and the result, by its own account, improved performance by more than 70 percent in eight weeks. The lab is aiming for a frontier-level EXAONE model by January 2027, and this free API is the bridge between the August release and that goal.
That framing is worth keeping in mind throughout this article. K-EXAONE 2.5 is not a finished consumer product with a polished app and a support desk. It is a hosted research-grade model, offered free, so that developers can test it in real work. For the right user — someone building agents, prototypes, or internal tools — that is exactly what makes it interesting. For someone looking for a polished chatbot to replace their current assistant, it is the wrong tool.
The Free API: What LG Actually Opened Up
The EXAONE API Platform is LG’s hosted service for its foundation models, and the free tier is the headline. According to LG’s announcement and the platform’s own changelog, the service is currently free of charge. Two models are available through it:
- K-EXAONE 2.5 — the flagship agentic language model, the subject of this article.
- EXAONE Tabular — a smaller model built for predictions on structured data, such as tables and spreadsheets.
Anyone can sign up for an API key at the EXAONE API Platform. LG’s announcement names developers, startups, companies, and research institutions explicitly, and says the model can be used for service development and research. There is no mention of a waitlist, an application process, or a credit-card requirement in the announcement — you get a key and you call the endpoint.
Two things LG has not said are just as important. First, the company has not said how long the free period will last. The platform changelog says the service is “currently free of charge,” which is the language of a launch offer, not a permanent price. Second, the announcement publishes no rate limits, quotas, or usage caps. That does not mean there are none — it means you should read the platform’s terms before you route production traffic through it, and you should not be surprised if limits appear later.
It is also worth being precise about what “free” covers. This is API access only. LG has published weights for earlier K-EXAONE models — version 2.0 was released under the Apache 2.0 open-weight licence — but the company has announced no open weights for version 2.5. You cannot download K-EXAONE 2.5 and run it on your own hardware. You call LG’s hosted endpoint, and LG runs the model. That distinction matters for privacy, for latency, and for anyone whose compliance rules forbid sending data to a third-party API.
K-EXAONE 2.5 Specs at a Glance
Here are the numbers LG publishes for the model, taken from the EXAONE API Platform documentation and the model card:
- Total parameters: 750 billion
- Active parameters per token: 31 billion (mixture-of-experts architecture — only a fraction of the model fires for each token, which is how a 750B model stays affordable to serve)
- Context window: 512,000 tokens
- Combined request cap: input tokens plus the generation allowance cannot exceed 524,288 tokens; larger requests return a 400 error
- Maximum output: 512,000 tokens
- Modalities: text only — no image input, no image output, no audio
- Languages: Korean and English
- Knowledge cutoff: January 2026
- Endpoint: OpenAI-compatible POST /v1/chat/completions, model identifier
k-exaone_v2.5 - Reasoning control: a
reasoning_effortparameter set to low, high, or max (max is the default), controlling how much the model thinks before answering - Price: free at launch; no paid tiers announced
The 512K context window deserves a pause, because it is the spec most likely to change how you use the model. Half a million tokens is roughly 350,000–400,000 words of English text — enough to paste in a long contract, a whole codebase’s worth of relevant files, or months of meeting notes and still have room for the answer. Most paid frontier APIs charge real money for that much context. Here it is free, at least for now.
The reasoning control is the other spec worth understanding. The reasoning_effort parameter lets you trade thinking time for speed and cost. Set it to low for quick, cheap answers to simple questions; leave it at the default max when you want the model to work through a hard problem step by step. That kind of dial is common on frontier APIs, and its presence here is a sign that LG is positioning K-EXAONE 2.5 as a serious tool for agent workloads, not a demo.

How the K-EXAONE 2.5 Free API Works
The most practical thing about the model is how little new code it asks of you. The endpoint is OpenAI-compatible, which means any client library, framework, or tool that already talks to an OpenAI-style chat completions endpoint can point at LG’s platform with two changes: a different base URL and a different model name.
The flow looks like this. You create an account on the EXAONE API Platform at docs.exaone.ai and generate an API key. You send a POST request to the platform’s /v1/chat/completions endpoint, with the model field set to k-exaone_v2.5 and your messages in the standard chat format. The platform returns the model’s reply, optionally streamed token by token. That is the whole integration for a basic chat call.
A minimal request, in the shape every OpenAI-compatible client already uses, looks like this:
POST {EXAONE_API_BASE_URL}/v1/chat/completions
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json
{
"model": "k-exaone_v2.5",
"messages": [
{"role": "user", "content": "Summarise this contract in plain English."}
],
"reasoning_effort": "high",
"stream": true
}
Check the exact base URL in the platform documentation before you write your adapter — LG’s newsroom announcement does not spell out the request format, and the docs are the authoritative source. The model identifier and the endpoint shape above come from the platform’s own endpoint reference, but base URLs and authentication details are the kind of thing that changes between a launch announcement and the live docs.
Beyond plain chat, the documentation lists four capabilities that matter for real work:
- Tool calling — the model can decide to call a function you define, receive the result, and continue. This is the foundation of agent workflows: the model does not just answer, it acts.
- Structured outputs — you can require the reply in a fixed format, such as JSON matching a schema you supply. That turns the model into a reliable component in a pipeline instead of a text box you have to parse by hand.
- Code generation — the model writes and explains code, which is where LG’s agentic benchmark claims are concentrated.
- Streaming — responses arrive token by token, so your interface can show progress instead of making the user wait for the whole answer.
Because the interface is OpenAI-compatible, the practical switching cost is low. If you already have a prototype running against another OpenAI-compatible endpoint, pointing it at LG’s endpoint is a configuration change, not a rewrite. That is a deliberate design choice by LG: the fastest way to get developers to try a new model is to make trying it boring.

K-EXAONE 2.5 Benchmarks: What LG Claims — and What It Does Not Prove
LG’s headline claim is that K-EXAONE 2.5 now matches DeepSeek V4.1 Flash and GLM-5.3 on agentic benchmarks, and that it is competitive with frontier-class open-weight models. The company says it measured the model internally on publicly available datasets, and reports that performance improved by more than 70 percent in the eight weeks since the August release.
Those numbers need careful handling, because every one of them is vendor-reported. LG published no sample counts, no decoding settings, no tool harnesses, and no description of the agentic scaffolds used in its tests. The scores have not been independently audited or reproduced by a third party. That does not make them false — it makes them unverified, and the honest way to report them is as LG’s own measurements, not as settled fact.
There are three specific limits to keep in mind when you read the benchmark story:
- They are self-evaluations. The lab that built the model also ran the tests and chose what to publish. Independent benchmarks, such as the public leaderboards that track other models, have not yet weighed in on K-EXAONE 2.5.
- The comparison set is narrow. “Competitive with frontier-class open-weight models” and “matches DeepSeek V4.1 Flash and GLM-5.3” are claims about a specific slice of models on a specific slice of tasks — agentic work, not general knowledge, not creative writing, not multilingual fluency.
- The baseline moved fast. A 70 percent improvement in eight weeks sounds dramatic, but it is measured against LG’s own August model, which was itself a new release. Rapid improvement from a low starting point is normal; it is not the same as a 70 percent lead over the field.
None of this is a reason to dismiss the model. It is a reason to test it yourself — which is, conveniently, free. The right response to a vendor benchmark is not belief or disbelief, but a small experiment: take a task you already do with another model, run it through the model, and compare the outputs yourself. The free API exists precisely so you can do that without paying for the privilege.
For context on the comparison LG itself invites, DeepSeek V4.1 Flash is the efficiency release DeepSeek shipped in October 2026 — a model our own coverage found beating Claude Sonnet 5.6 on most coding benchmarks in its launch reporting, while producing roughly 20 percent fewer tokens. If it genuinely sits in that class on agentic tasks, it is a serious model. The only way to know is to run it. You can read our DeepSeek V4.1 Flash explainer for the full picture of the model LG is measuring itself against.
What You Can Build With K-EXAONE 2.5
The free API is most useful where its specific strengths line up with the job. Here are the use cases that make sense, based on what LG actually ships:
Agentic coding assistants. This is the headline use case, and the one LG’s benchmarks target. A coding agent built on the model can read a repository, plan a change, call tools to inspect files and run tests, and iterate. The 512K context window helps here more than anywhere else: the agent can hold far more of your codebase in view at once, which means fewer “I can’t see that file” failures. The reasoning_effort dial lets you spend more thinking time on hard refactors and less on trivial edits.
Document work in Korean and English. The model’s bilingual strength is a genuine differentiator. If your team works across Korean and English documents — contracts, research papers, technical specs, support tickets — the model can summarise, translate, extract structured data, and answer questions across both languages in a single call. Long documents are the natural fit: the 512K window swallows contracts and reports whole, instead of forcing you to chunk them and stitch the answers together.
Structured data extraction. With structured outputs, you can ask the model to read messy text — emails, PDFs, support threads — and return clean JSON matching your schema. Invoices, purchase orders, meeting notes, and research abstracts all become database rows. Because the API is free, this is a cheap way to prototype an extraction pipeline before you commit to a paid provider.
Internal tools and prototypes. The combination of “free” and “OpenAI-compatible” makes the model an excellent prototyping backend. Build the internal dashboard, the Slack bot, the report generator, or the customer-support draft tool against the free endpoint. If the prototype proves its value, you can decide later whether to stay, switch to a paid model, or self-host something else — the OpenAI-compatible interface means you are not locked in.
Research and evaluation. For researchers and students, a free 750B model is a chance to study how a frontier-class model behaves on your own tasks, without a cloud bill. That includes benchmarking it against other models, testing its Korean-language ability, and probing where it fails. LG explicitly invites research use, and the free API is the mechanism.
What the model is not built for is equally clear. It is text-only, so any workflow involving images, audio, or video needs another model. It knows Korean and English, so other languages are out of scope. And because it is a hosted API with no published data-retention terms in the announcement, it is not the right choice for sensitive data until you have read the platform’s terms and are comfortable with them.

K-EXAONE 2.5 vs K-EXAONE 2.0: What Changed
To understand version 2.5, it helps to know what version 2.0 was. K-EXAONE 2.0, released in August 2026, was the second model in LG’s sovereign AI line. It was an open-weight release under the Apache 2.0 licence, built on a mixture-of-experts architecture with 37 billion active parameters and a 262,144-token context window. Developers could download the weights and run the model themselves.
K-EXAONE 2.5 is a different kind of release. LG applied additional training to the August model and now serves the result through its own hosted platform. The headline numbers moved in the directions you would expect from three more months of training: the context window roughly doubled, from 262K to 512K tokens, and the active parameter count stepped down slightly, from 37 billion to 31 billion, while total parameters reached 750 billion. LG reports the performance gain at more than 70 percent over eight weeks — its own measurement, on its own tests, as discussed above.
But the bigger change is the distribution model. Version 2.0 was weights you could download. Version 2.5 is an API you call. That shift tells you something about LG’s strategy: the lab wants usage data and real-service feedback, and a hosted API gives it both, while an open-weight release gives it neither. It also tells you something about the model’s status. LG calls 2.5 a technology-validation model — a checkpoint on the road to the frontier-level EXAONE model it aims to complete by January 2027 — not a finished product to be frozen and distributed.
For users, the practical difference is simple. If you used K-EXAONE 2.0, you could self-host it, fine-tune it, and run it offline. With version 2.5, you get a stronger model and a longer context window, but you give up self-hosting, you send your prompts to LG’s servers, and you accept that the free tier could change. Whether that trade is worth it depends on what you are building. For most prototyping and evaluation work, the stronger hosted model wins. For privacy-sensitive or offline work, the older open-weight release — or another open model entirely — is still the better fit.
The Sovereign AI Story Behind K-EXAONE 2.5
K-EXAONE 2.5 is not just a product launch. It is a piece of national industrial policy, and understanding that context explains several of the choices LG made — including the free API itself.
South Korea’s sovereign AI foundation-model project exists to build homegrown AI models that reduce the country’s reliance on foreign systems. The government backing means the project is judged partly on capability and partly on adoption: a Korean model that nobody uses does not serve the goal, however good its benchmarks are. Opening the API for free is the fastest way to get real developers using the model in real services, which generates the usage data and the feedback loop the project needs.
The launch venue reinforces the point. LG announced the free API at NVIDIA AI Day Seoul 2026, a showcase event for Korea’s AI ecosystem, with Nvidia’s hardware story in the background. LG is also considering Nvidia’s Dynamo technology to improve inference efficiency and service scalability — a possible infrastructure step, not a confirmed deployment, but a sign of how the lab thinks about serving a 750B model at scale.
There is a broader pattern here that US readers should recognise. Sovereign AI programmes are now running in South Korea, in the European Union, in India, and elsewhere, and each one produces models that eventually reach global developers. The practical consequence is a wider, more diverse menu of models than the US–China duopoly of a few years ago. For a developer, that is good news: more competition means more free tiers, more pricing pressure, and more choice. K-EXAONE 2.5 is the latest entry on that menu, and its free API is the easiest way to taste it.
None of this changes the technical facts. A model’s origin story does not make its answers better or worse. But it does explain why LG is giving away access to a 750-billion-parameter model, and it sets a reasonable expectation: the free tier is a strategic move, not a charity, and strategies change. Use it while it is free, and build in a way that survives it becoming paid.
K-EXAONE 2.5 vs DeepSeek V4.1 Flash and Other Free Options
LG invites the comparison with DeepSeek V4.1 Flash and GLM-5.3, so it is worth laying out what that comparison actually means — and where it stops.
Against DeepSeek V4.1 Flash. DeepSeek’s October 2026 efficiency release is the model LG says K-EXAONE 2.5 now matches on agentic benchmarks. DeepSeek’s model is a paid API product — our own coverage of the launch reported input pricing around $0.08 per million tokens for the Flash variant, with the flagship V4 Pro higher — and it is known for producing roughly 20 percent fewer tokens than Claude Sonnet 5.6 at the same intelligence level on coding benchmarks. K-EXAONE 2.5 answers that with a free API and a 512K context window. On price alone, LG wins the comparison for now. On proven, independently measured coding performance, DeepSeek has the public track record and LG does not — yet.
Against GLM-5.3. GLM is the Chinese frontier model LG also names as a benchmark peer. It sits in the same class: a large, capable model aimed at agent workloads. The practical difference for a US developer is access and language. K-EXAONE 2.5 offers a free, English-capable, OpenAI-compatible endpoint today; GLM’s availability and pricing vary by channel. If your work is in English and you want to start in ten minutes, the LG option is the simpler path.
Against the open-weight field. This is where the comparison gets awkward for LG. Models like Meta’s Muse Spark Open and Liquid AI’s efficient open-weight releases can be downloaded and run on your own hardware, which K-EXAONE 2.5 cannot — the 2.5 release is API-only. If self-hosting, offline use, or fine-tuning matters to you, the open-weight field still wins, whatever LG’s benchmark scores say. You can read our Meta Muse Spark Open explainer for the strongest current example of that approach. The honest summary is that the model competes on hosted capability and price, not on openness.
The bottom line of the comparison: on the two dimensions that matter most for trying a new model — cost and ease of access — K-EXAONE 2.5 is currently the most generous offer among frontier-class models. On independently verified performance, it is still an open question. That combination — cheap to try, unproven at scale — is exactly why the right move is to test it on your own work rather than to take anyone’s word for it, including ours.
Honest Limitations to Know Before You Commit
A free 750B model sounds too good to be true, and in the usual sense it is: the offer is real, but it comes with limits that matter. Here they are, plainly stated.
The free period has no published end date — and no published permanence either. LG says the service is currently free. It has not said the free tier is permanent, and it has not said what happens when it ends. Build anything important on this API with a migration plan in mind, and keep your integration behind the OpenAI-compatible interface so switching providers is a configuration change.
There are no published rate limits or quotas. The announcement does not state how many requests you can make, how many tokens you can use, or what happens when you hit a limit. In practice, free tiers almost always have limits, and they usually appear in the platform terms rather than the press release. Read the terms at docs.exaone.ai before you depend on the API.
It is text-only, in Korean and English. No images, no audio, no video, and no other languages. If your workflow needs vision, speech, or a third language, you will need a second model alongside K-EXAONE 2.5.
The benchmarks are LG’s own. As covered above, the agentic scores are internal, unaudited, and published without the details that would let anyone reproduce them. Treat them as a reason to test the model, not as proof of its quality.
There are no open weights for 2.5. You cannot download it, fine-tune it, or run it offline. Your prompts go to LG’s servers. The announcement also publishes no data-retention or privacy terms, so if you handle sensitive data, you need to read the platform’s terms and decide for yourself whether the arrangement is acceptable.
The knowledge cutoff is January 2026. The model does not know about events after that date unless you put them in the prompt. For a model launched in November 2026, that is a ten-month gap. It matters less for coding and document work, and more for anything involving current events, recent product changes, or new regulations.
The 512K context is a ceiling, not a guarantee. The combined input-plus-output cap is 524,288 tokens, and very long contexts can still degrade answer quality on any model. Long context is a tool, not a magic trick: the model still has to find the right passage in a haystack of your own making.

How to Start Using the K-EXAONE 2.5 Free API in 10 Minutes
Here is the fastest honest path from zero to a working call.
Step 1: Get a key. Go to docs.exaone.ai, create an account on the EXAONE API Platform, and generate an API key. Keep the key somewhere safe — treat it like a password, because anyone with it can spend your quota.
Step 2: Read the terms. Before you send anything real, read the platform’s terms on retention, permitted use, region, and quota. This is the step most people skip and most people later regret skipping. It takes five minutes and it answers the questions the announcement does not: what LG does with your prompts, what counts as allowed use, and what the limits are.
Step 3: Make a test call. Send a simple chat completion with the model set to k-exaone_v2.5. Use reasoning_effort set to low for the first call, so you get a fast answer and confirm the plumbing works. If you use an OpenAI-compatible client library, this is a base-URL and model-name change.
Step 4: Run your own benchmark. Pick a task you already do with another model — a code review, a document summary, a structured extraction — and run the same prompt through K-EXAONE 2.5 and your current model. Compare the outputs side by side. Pay attention to the things benchmarks do not measure: whether the answer follows your format, whether the code actually runs, whether the summary drops anything important.
Step 5: Try the agentic features. Once the basics work, test tool calling and structured outputs on a small, safe task — for example, a function that looks up a value from your own data and returns it in JSON. This is where K-EXAONE 2.5 claims its strength, so this is where you should verify it.
Step 6: Decide what it is for. If the model does your job well, wire it into the prototype or internal tool you had in mind — behind an interface you can swap later. If it does not, you have spent ten minutes and no money finding out, which is the whole point of a free tier.
Who K-EXAONE 2.5 Is Best For — and Who Should Skip It
K-EXAONE 2.5 is best for developers and small teams who want to experiment with a frontier-class model without a cloud bill. If you are building an agent prototype, testing whether a large model improves your document pipeline, or evaluating models for a project that has not chosen its backend yet, the free API removes the usual cost barrier to a serious test. It is also a good fit for researchers, students, and anyone working in Korean and English who wants a bilingual model with a very long context window.
It is a particularly good fit if you already work with OpenAI-compatible tooling. The switching cost is near zero, and you can run the same prompts against K-EXAONE 2.5 and your current provider and compare the results directly. That kind of head-to-head test is the most valuable thing a free tier can offer, and this one is unusually generous.
Who should skip it? Anyone who needs image or audio input, anyone working in languages other than Korean and English, and anyone who needs to self-host or fine-tune — the API-only release cannot do any of those things. Anyone handling sensitive or regulated data should also pause until the platform’s data-retention terms are clear. And anyone building a production service that must keep running next year should treat the free tier as a trial, not a foundation: the price is “currently free,” not “free forever,” and a production dependency on an unpriced tier is a risk, not a saving.
There is also a security angle worth noting for the technically minded. The same week this free API launched, the industry was still digesting the news that Claude had helped uncover critical Xen hypervisor flaws — a reminder that AI models are now part of the security landscape, not just the productivity landscape. You can read our coverage of the Claude Xen vulnerability for that story. The connection here is practical: if you build agents on any model, including K-EXAONE 2.5, give them the least privilege they need, and never let an agent act on a system it should only be reading.
The Verdict
K-EXAONE 2.5 is the most generous free model offer of late 2026: a 750-billion-parameter model, a 512,000-token context window, an OpenAI-compatible endpoint, and a zero-dollar price tag. LG’s benchmark claims are unverified, the free period has no published end, and the model is text-only in two languages — but none of those caveats costs you anything to check, because the checking itself is free.
The right way to think about this release is as a ten-minute experiment, not a commitment. Get a key, run your own task through it, and compare the result with whatever you use today. If it holds up on your work, you have found a free backend for your prototype. If it does not, you have spent ten minutes learning that, which is a good trade either way.
What is genuinely new here is not the model alone. It is the pattern: a sovereign AI programme, built with government backing, giving its flagship model away for free to win developers. That pattern is spreading, and it is good for everyone who builds with AI. More models, more free tiers, and more competition mean better tools at lower prices. K-EXAONE 2.5 is the latest proof of that — and it is free to try today.
FAQs About K-EXAONE 2.5
Is the K-EXAONE 2.5 API really free?
Yes, for now. LG AI Research announced on November 9, 2026, that the API is free of charge through the EXAONE API Platform, and the platform’s changelog confirms the service is currently free. LG has not said how long the free period will last or what pricing will follow, so treat it as a launch offer rather than a permanent price.
How do I get a K-EXAONE 2.5 API key?
Create an account on the EXAONE API Platform at docs.exaone.ai and generate an API key. The announcement says developers, startups, companies, and research institutions can all obtain a key for service development and research.
What is the K-EXAONE 2.5 model identifier?
The model identifier is k-exaone_v2.5, used with an OpenAI-compatible POST /v1/chat/completions endpoint. Check the platform documentation for the exact base URL before you write your integration.
Is K-EXAONE 2.5 open source?
No — not version 2.5. The earlier K-EXAONE 2.0 was released under the Apache 2.0 open-weight licence, but LG has announced no open weights for version 2.5. This release is API access only, through LG’s hosted platform.
What can K-EXAONE 2.5 do?
It is a text-only model for Korean and English, with chat, tool calling, structured outputs, code generation, and streaming. It has a 512,000-token context window and a reasoning-effort control (low, high, or max). It cannot process images, audio, or video, and it does not work in other languages.
How good is K-EXAONE 2.5 compared with DeepSeek V4.1 Flash?
LG says its internal measurements show the model matching DeepSeek V4.1 Flash and GLM-5.3 on agentic benchmarks. Those scores are vendor-reported and unaudited — no independent benchmark has confirmed them yet. The reliable way to compare is to run your own tasks through both models, and the free API makes that easy.
What is the context limit for K-EXAONE 2.5?
The context window is 512,000 tokens. Input tokens plus the generation allowance cannot exceed 524,288 tokens in a single request; larger requests return a 400 error. The maximum output is 512,000 tokens.
Can I use K-EXAONE 2.5 for commercial services?
LG says the free API can be used for service development and research, and names developers, startups, companies, and research institutions as eligible users. Read the EXAONE API Platform’s terms for the exact permitted-use rules before you ship anything commercial on it.
