LG AI Research Opens Free K-EXAONE 2.5 API: 750B Total, 31B Active, 512K Context, with a Self-Reported 59.2 on SWE-Bench Pro, Just Above DeepSeek-V4.1-Flash

Background

Under Korea's sovereign AI foundation model program, LG AI Research released the open-weight K-EXAONE 2.0 in August. On November 9 at NVIDIA AI Day Seoul at COEX, LG made K-EXAONE 2.5, a further-trained version of 2.0, free to use through its newly launched EXAONE API Platform. LG calls it a technical validation model, and the training methods it validates will feed into the third-round model. According to the Korea Times, LG aims to deliver a frontier-level EXAONE model by January.

Model and API

Per the official model page: 750B total parameters with 31B active (2.0 had 37B active), text-only, Korean and English, a 512K context and 512K max output, and a January 2026 knowledge cutoff. The API is OpenAI Chat Completions compatible with model name k-exaone_v2.5. The model always reasons before answering; reasoning_effort takes low, high or max (default max), and the reasoning comes back separately in reasoning_content. Input tokens plus max_tokens may not exceed 524,288, or the API returns a 400 error. NVIDIA's blog says the model was trained with NVIDIA Nemotron datasets, including Nemotron-Personas-Korea, a set of 7 million synthetic Korean personas.

Vendor-reported benchmarks

LG ran internal evaluations on public datasets across four agentic benchmarks: SWE-Bench Pro 59.21, BrowseComp 78, AutomationBench (open set) 74.6 and Claw-Eval 77.29. Against DeepSeek-V4.1-Flash (58.41 / 75.8 / 74.9 / 75.78) and GLM-5.3 (64.03 / 79 / 71.6 / 77.64), results are mixed, and K-EXAONE 2.5 is not the top score among the three on any of them. Compared with 2.0, AutomationBench jumped from 16.9 to 74.6.

What it means for developers

For teams that need Korean and want to try a large MoE model at no cost, this is a new option that plugs into existing OpenAI SDKs and coding harnesses. Caveats: LG has not published the free period, quotas or rate limits, has not said whether 2.5 weights will be released, and the model does not list Chinese support. The default max reasoning adds latency and tokens, so tune reasoning_effort per task and enable preserve_thinking for multi-turn agent runs. All benchmark numbers are LG's own, so re-test on your own workloads before production use.