{"id":"k-exaone-2-5","slug":"k-exaone-2-5","name":"K-EXAONE 2.5","orgId":"lg","family":"k-exaone-2-5","releaseDate":null,"availability":"ga","openWeights":false,"license":null,"paramsTotal":750000000000,"paramsSource":"tracked","contextWindow":null,"knowledgeCutoff":null,"modelCardUrl":"https://www.lgresearch.ai/news/view?seq=755","hfRepo":null,"modalities":["text"],"aliases":["K-EXAONE 2.5","k-exaone 2.5","k-exaone-2.5","k-exaone-2-5","k-exaone_v2.5","K-EXAONE-2.5","K EXAONE 2.5","EXAONE 2.5","LG K-EXAONE 2.5","LG AI Research K-EXAONE 2.5"],"aliasNote":null,"origin":"tracked","tracked":{"trackedSince":"2026-11-14","announced":"2026-11-09","blurb":"LG AI Research's text language model with 750B parameters (31B active). Official API docs say it is OpenAI-compatible, text-only, and has a 524,288-token context window (512K on the official models page), with a January 2026 knowledge cutoff. Official pages describe it as a technical validation model given additional training from the August 2026 K-EXAONE 2.0 release, with strengths in Korean and English, coding, long-context tool calling, and agent tasks. It is available for free through the public EXAONE API Platform as model k-exaone_v2.5. Official Terms of Use, effective November 9, 2026, grant a non-commercial right to use the platform and state that weights are not provided.","api":"live","apiModelId":"k-exaone_v2.5","openWeights":"no","params":{"text":"750B (31B active)","source":"Official EXAONE API Platform models page","total":750000000000},"facts":[{"label":"Architecture","value":"Text-only language model. Official models page: 750B parameters (31B active); languages Korean and English; features chat, tool calling, structured outputs, code generation, and streaming. Official 2.5 pages do not describe a mixture-of-experts layout beyond the 31B active figure"},{"label":"Context","value":"Official endpoint docs: maximum context length is 524,288 tokens (input tokens plus max_tokens). Official models page lists context length 512K and max output tokens 512K"},{"label":"Knowledge cutoff","value":"2026-01 (official models page)"},{"label":"API","value":"Free public OpenAI-compatible Chat Completions API at https://api.exaone.ai/v1/chat/completions. Official model name: k-exaone_v2.5. Official changelog, November 9, 2026: EXAONE foundation models are available through the API and the service is currently free of charge. Official news: developers, startups, enterprises, research institutions, and anyone else can use it at no cost by obtaining an API key. Official @LG_AI_Research post: API access is open to everybody, with no approval needed"},{"label":"Terms","value":"Official EXAONE API Platform Terms of Use, effective November 9, 2026: a non-commercial right to use the platform; commercial use of the model or output, and integrating the API into products or services so a third party can call the model, are not permitted. The terms also state that weights, parameters, and source code are not provided. API keys expire 30 days after issuance. Users must be 19 or older"},{"label":"Weights","value":"No public weight release for K-EXAONE 2.5 is stated on official LG AI Research news, the EXAONE API docs, or the official Terms, which say weights are not provided. Official Hugging Face pages under LGAI-EXAONE for guessed 2.5 repo names return 404. Official news describes K-EXAONE 2.5 as LG AI Research's proprietary AI foundation model"},{"label":"Reasoning","value":"Official endpoint docs: the model always reasons before answering, and enable_thinking cannot be disabled. reasoning_effort may be low, high, or max (default max). Reasoning is returned in message.reasoning_content and counted in usage.reasoning_tokens"},{"label":"Predecessor","value":"Official news: a technical validation model that received additional training based on the version released in August 2026. Official models page compares it with K-EXAONE 2.0"},{"label":"Training","value":"Official models page: from continual pre-training through post-training, a high-quality synthetic data methodology was applied, and an anomalous data detection system was used to curate and refine the training data"}],"claims":["Lab claim: self-reported on the official EXAONE API Platform models page, SWE-Bench Pro 59.21, BrowseComp 78, AutomationBench (Open-set Evaluation) 74.6, and Claw-Eval 77.29. Official November 9, 2026 news describes these as internal evaluations using publicly available datasets and rounds SWE-Bench Pro to 59.2 and Claw-Eval to 77.3.","Lab claim: self-reported on the same official models table, predecessor K-EXAONE 2.0 scores of SWE-Bench Pro 47.1, BrowseComp 50.87, AutomationBench (Open-set Evaluation) 16.9, and Claw-Eval 62.11. The table also lists comparison scores for DeepSeek-V4.1-Flash and GLM-5.3."],"sources":[{"kind":"lab","title":"K-EXAONE 2.5 API Available for Free","url":"https://www.lgresearch.ai/news/view?seq=755","date":"2026-11-09"},{"kind":"lab","title":"K-EXAONE 2.5 API 무료 공개","url":"https://www.lgresearch.ai/news/view?seq=754","date":"2026-11-09"},{"kind":"lab","title":"K-EXAONE 2.5 API docs","url":"https://docs.exaone.ai/en/docs/k-exaone-2.5","date":"2026-11-09"},{"kind":"lab","title":"EXAONE API Platform models","url":"https://docs.exaone.ai/en/docs/models"},{"kind":"lab","title":"EXAONE API Platform changelog","url":"https://docs.exaone.ai/en/docs/changelog","date":"2026-11-09"},{"kind":"lab","title":"EXAONE API Platform Terms of Use","url":"https://docs.exaone.ai/en/terms/terms-of-service","date":"2026-11-09"},{"kind":"lab","title":"LG AI Research announcement on X","url":"https://x.com/LG_AI_Research/status/2119698881961497062","date":"2026-11-09"},{"kind":"lab","title":"LG AI Research API access post on X","url":"https://x.com/LG_AI_Research/status/2119700328883114096","date":"2026-11-09"}],"firstBenchmark":null},"organization":{"id":"lg","name":"LG","slug":"lg","country":null,"url":"https://www.lgresearch.ai","color":"#A39E8F","mark":"LG"},"price":null,"answer":{"word":"Pending benchmarks.","scope":"Announced Nov 9, 2026; no independent results yet.","lead":"As of Nov 22, 2026, K-EXAONE 2.5 is pending benchmarks on isitsota: announced Nov 9, 2026, with no independent results on the boards we track yet.","confidence":"low"},"status":"PENDING_BENCHMARKS","verdicts":[]}