LG's K-EXAONE 2.5 Free API: A 750B Agent Model You Can Try Today
Amir Farid
7 min read
LG AI Research has opened free API access to K-EXAONE 2.5, a 750-billion-parameter model built for agentic coding and workplace automation. Announced on 9 November 2026, it is available as a hosted API through the EXAONE API Platform: developers, startups and companies can obtain an API key and use it for service development and research. It is text-only, works in Korean and English, and uses the same OpenAI-compatible chat completions interface most teams already use. In short, you can test a frontier-class coding model in your existing tooling today, free.
What LG actually announced
The announcement, covered in Unite.AI's report on the launch, is deliberately narrow. LG is not selling a product or a platform, just hosted access to a model it calls a technology-validation model, trained further from the August 2026 version. You get an API key, you call the model, you build with it.
A technology-validation model is a working model in developers' hands. The earlier K-EXAONE 2.0 was an open-weight Apache 2.0 release: 37 billion active parameters, 262,144-token context. K-EXAONE 2.5 is a different offer: the weights are not open. What is free is hosted API access.
The model in numbers
K-EXAONE 2.5 is a 750-billion-parameter model with 31 billion active parameters — a mixture-of-experts design where only a fraction runs on any given token. The documented facts:
- 750 billion total parameters, 31 billion active (2.0: 37 billion active).
- Text-only. No image, audio or video input.
- Korean and English — two languages, with Korean a genuine strength.
- January 2026 knowledge cutoff. Anything later is outside its training.
- 524,288-token ceiling. Input plus max_tokens — reasoning included — cannot exceed 524,288, or the request returns a 400 error.
LG's benchmark scores — read them as LG's claims
LG says it measured K-EXAONE 2.5 internally on open datasets and found it competitive with frontier-class open-weight models. These are LG's own measurements, not independent verification. LG names code generation, debugging, refactoring and test-code writing as the core uses, alongside agentic search and workplace automation:
- SWE-Bench Pro, agentic coding: 59.21 (2.0: 47.1; DeepSeek-V4.1-Flash: 58.41; GLM-5.3: 64.03).
- BrowseComp, agentic search: 78.0 (2.0: 50.87; DeepSeek-V4.1-Flash: 75.8; GLM-5.3: 79.0).
- AutomationBench open-set, agentic tool use: 74.6 (2.0: 16.9; DeepSeek-V4.1-Flash: 74.9; GLM-5.3: 71.6).
- Claw-Eval, agentic tool use: 77.29 (2.0: 62.11; DeepSeek-V4.1-Flash: 75.78; GLM-5.3: 77.64).
The jump over K-EXAONE 2.0 is large on every benchmark — most dramatically AutomationBench, from 16.9 to 74.6. Against the two external models LG chose, it is competitive but not dominant: ahead of DeepSeek-V4.1-Flash on three of four, ahead of GLM-5.3 on only one. The only score that counts is the one you measure on your own tasks.
How the API actually works
The endpoint is an OpenAI-compatible POST /v1/chat/completions call with model id k-exaone_v2.5, returning standard OpenAI-format JSON or streamed chunks. If your tooling already talks to an OpenAI-compatible API, swapping the base URL and model id is a configuration change, not a rebuild. Teams building custom software can compare it against their current model on identical prompts.
Two controls matter. First, reasoning_effort defaults to max, and higher reasoning intensity can increase response time and token usage. Second, preserve_thinking carries earlier reasoning into multi-turn requests.
Hermes and OpenCode configs
LG's agent guide includes configurations for the Hermes general agent and the OpenCode coding agent. Your harness executes the tool calls the model returns and passes results back — so sandbox execution, cap the loop, log every action, keep a human in the loop. Our comparison of AI chatbots versus AI agents draws the line between answering and acting.
What to try first
Free access only helps if the trial answers a real question. Run one task you already do with a paid model on both, and score the results blind:
- Test-code writing. Give it a module and ask for edge-case tests — they pass or they do not.
- Debugging a real bug. Hand it a failing test and the relevant files: cause or symptom?
- Refactoring a messy function. Ask for a refactor with the tests as the safety net; judge it on preserved behaviour, not an elegant diff.
- Agentic search. Ask it to find a fact in your documentation and cite its sources — the retrieval problem our RAG chatbots for business guide explains.
- A small automation. A routine task across your SaaS tools, run through an agent harness. Measure correct completions, not fluent prose — see our guide to AI agents for small business.
Keep it small and honest: one workflow, one week, real tasks, a human reviewing every output — the same discipline we recommend for a voice agent. For help designing the comparison, talk to Coheser about an AI agent pilot.
What to watch before you build on it
Free is not free of risk. Text-only, two languages: no images, no voice, no documents beyond text. A support agent that reads screenshots, or a WhatsApp bot handling voice notes, needs a second model — and if your customers write in Urdu, Arabic or French, this model is not for them.
Reasoning costs time and tokens: with reasoning_effort defaulting to max, tune the effort per task — reasoning tokens count toward the 524,288 ceiling. LG's scores are self-measured: the benchmarks are LG's own internal runs on open datasets, and the January 2026 cutoff means newer libraries need to be supplied in context — the same caution that applies in regulated settings, as our guide to AI chatbots for clinics shows.
It is a validation model, not a product promise: free access today does not guarantee free access tomorrow. Build your agent so the model is swappable — the OpenAI-compatible interface makes that easy — rather than wiring your product to one free tier, the same portability argument our n8n versus Make versus Zapier comparison makes.
Free API versus paid frontier models — when it fits
The honest comparison is not K-EXAONE 2.5 against the best paid models on raw capability: on LG's own numbers it sits just behind GLM-5.3 on most benchmarks. What matters is cost of experimentation — a free, OpenAI-compatible model you can test this afternoon, against a paid model whose bill arrives with every experiment. For prototyping and Korean-language coding work, the free API deserves the test; for production, paid models bring service guarantees and a budgetable price, which a free validation API lacks. Prototype free, then choose the production model on measured results — the approach our guide to AI automation for logistics companies recommends, with voice alongside text in a full customer stack: see our AI calling agent.
What your business should do now
Start this week, while access is free. Get an API key from the EXAONE API Platform, point your existing agent harness at k-exaone_v2.5, and run the five trials above on real work. Score them honestly, with a human reviewing every output. If it beats your current model, you have found a cheaper brain for your agent; if not, you have spent an afternoon learning that, not a month's budget. Coheser builds AI agents and automation on that foundation, and our guides to custom software costs in the UK and scaling a SaaS MVP without rebuilding set out what the surrounding product costs once the model is free.
Frequently asked questions
Is the K-EXAONE 2.5 API really free?
LG AI Research has opened free API access through the EXAONE API Platform: you obtain an API key and use the model for service development and research. No pricing for continued use was announced, so treat it as a testing offer and check the terms first.
Is K-EXAONE 2.5 open source?
No. Only the hosted API is free. The earlier K-EXAONE 2.0 was an open-weight Apache 2.0 release; K-EXAONE 2.5 is hosted only, and its weights have not been released.
What languages does K-EXAONE 2.5 support?
Korean and English, text only. There is no image, audio or video input, and no other languages are supported.
How do I call the K-EXAONE 2.5 API?
Through an OpenAI-compatible POST /v1/chat/completions endpoint, model id k-exaone_v2.5. Input plus max_tokens, including reasoning, cannot exceed 524,288. Reasoning effort defaults to max; preserve_thinking carries earlier reasoning into multi-turn requests.
Should my business switch to K-EXAONE 2.5?
Not on the announcement alone. LG's figures are its own internal measurements, not independently verified. Run a one-week pilot on your own tasks, score it against your current model, and decide on measured results.