LG AI Research is making the API for K-EXAONE 2.5 available free of charge. In a 9 November news post, the lab said developers, startups, enterprises and research institutions can use the model for service development and research at no cost after obtaining an API key on the EXAONE API Platform.
K-EXAONE 2.5 is described as a technical validation model that received additional training on top of a version released in August. The story is less a sudden model drop than a decision to widen access to that validation build while LG publishes its own agentic-performance numbers against other frontier-class open-weight systems.
Free access sits on an OpenAI-compatible API
According to the K-EXAONE 2.5 documentation, the model is text-only, uses the model id k-exaone_v2.5, and is reached through an OpenAI-compatible chat-completions endpoint. Maximum context length is 524,288 tokens; if the sum of input tokens and max_tokens exceeds that limit, the platform returns a 400 error.
The docs also expose a reasoning_effort setting of low, high or max, with max as the default. Reasoning content and the final answer are returned separately. LG’s news post points readers to docs.exaone.ai for the free API path and does not, in that notice, publish a public rate-limit schedule; the concrete invitation is free use with an API key, not an unbounded capacity claim.
That access pattern sits in a wider Asian open-model lane already covered by TSN, from Kakao’s Kanana-2 embedding releases to Baidu’s ERNIE-5.5-Flash-Base open weights. It also connects to broader sovereign AI infrastructure debates, where governments and large firms treat national model capacity as strategic infrastructure rather than a pure product launch.
LG’s agentic scores are internal, and mixed against GLM-5.3
LG says K-EXAONE 2.5 shows competitive agentic performance against frontier-class models in internal evaluations on public datasets. The company highlights four scores from those internal runs:
| Benchmark | LG-reported K-EXAONE 2.5 score | What LG says it measures |
|---|---|---|
| AutomationBench-Openset | 74.6 | Automating tasks with SaaS tools |
| SWE-Bench Pro | 59.2 | Real-world software development and modification |
| BrowseComp | 78.0 | Navigating the web to find information |
| Claw-Eval | 77.3 | Workplace problem-solving for agents |
LG compares those results with DeepSeek-V4.1-Flash and GLM-5.3. Press reports of LG’s comparison table put DeepSeek at 74.9 / 58.4 / 75.8 / 75.8 and GLM-5.3 at 71.6 / 64.0 / 79.0 / 77.6 on the same four benchmarks in that order. On that company-framed comparison, K-EXAONE 2.5 leads some columns and trails GLM-5.3 on SWE-Bench Pro and BrowseComp. None of these figures is an independent third-party leaderboard result.
LG also says the model can help with coding assistance, debugging, refactoring and test creation. Those are product claims from the same news post, not separate audited productivity studies.
Company scores are a starting map, not a league table
A free API changes who can try a model far more quickly than a weights dump changes who can host one. For startups and labs that mainly need an OpenAI-compatible endpoint, that is a practical difference: integration cost falls even when self-hosting remains out of reach. The 524,288-token window matters for long agent traces, but only if the caller stays inside the documented token budget and accepts that reasoning tokens count toward usage.
Internal benchmarks are useful as a self-portrait. They are less useful as a championship claim. When the same comparison set shows K-EXAONE 2.5 behind GLM-5.3 on software-engineering and browsing tests, the honest reading is “competitive on LG’s map,” not “the leader.” Readers comparing Korean, Chinese and Western agent stacks should treat each vendor’s sheet as advocacy data until the same harness is run by someone without a product to ship.
The next evidence that would move the story is operational rather than rhetorical: published rate limits or quotas, independent reproductions of the four agentic scores, and whether later K-EXAONE builds keep the free-API path or narrow it once validation ends. Until then, the durable fact is access plus attribution—free keys to a text-only validation model, with LG’s numbers labelled as LG’s.
Sources
- LG AI Research: K-EXAONE 2.5 API available for free, 9 November 2026
- EXAONE API Platform docs: K-EXAONE 2.5
- LG Corp media release on K-EXAONE 2.5 (Korean)
Related stories
Kakao’s Kanana-2 embedding models is the nearest Korean open-model counterpart: another domestic lab widening developer access, this time with embedding weights rather than a free chat API.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
