K-EXAONE 2.5 opens its free APIs to developers
LG AI Research makes K-EXAONE 2.5 freely accessible via the EXAONE API Platform. The 750-billion-parameter model targets agent use cases, from code to automation, with internal performance yet to be confirmed in production.
LG AI Research announced on November 9, 2026, free access to the K-EXAONE 2.5 API, a 750-billion-parameter model offered via the EXAONE API Platform. Developers, startups, companies, and research centers can request a key to use it for service development or research. This gesture matters as it lowers the entry cost to test a model oriented towards agents, code, and automation in tools already compatible with OpenAI.
A model designed for agents, not just chatting
LG presents K-EXAONE 2.5 as a technology validation model, obtained after additional training applied to the version released in August 2026. The model is text-only, supports Korean and English, and exposes functions that speak directly to product teams: chat, tool calls, structured outputs, code generation, and streaming. In other words, it does not just respond; it can also integrate into a workflow with actions to trigger around it.
The API documentation adds an OpenAI-compatible interface via a POST /v1/chat/completions with the identifier k-exaone_v2.5. Inputs plus maximum generation, including reasoning, cannot exceed 524,288 tokens; beyond that, the request returns a 400 error. That’s large but not magic: context still needs careful management, which often remains the real challenge in production.
Flattering internal scores, but a test ground yet to be confirmed
LG says it evaluated K-EXAONE 2.5 internally on open datasets and presents it as competitive against leading open-weight models. On published figures, it improves over K-EXAONE 2.0 on the four cited tests: 59.21 versus 47.1 on SWE-Bench Pro, 78.0 versus 50.87 on BrowseComp, 74.6 versus 16.9 on AutomationBench open-set, and 77.29 versus 62.11 on Claw-Eval. These benchmarks cover code, web information retrieval, automation with SaaS tools, and problem-solving in work environments.
Compared to DeepSeek-V4.1-Flash, K-EXAONE 2.5 performs better on SWE-Bench Pro, BrowseComp, and Claw-Eval, but lags on AutomationBench. Compared to GLM-5.3, it only leads on AutomationBench. The picture is therefore more nuanced than a simple “everything is faster”: depending on the use case, the advantage shifts.
The real test will be in real agents
LG names software development, debugging, refactoring, and test writing among the targeted uses, along with agentic research and work automation. The documentation also exposes reasoning controls: reasoning_effort is set by default to max, while preserve_thinking can pass prior reasoning in multi-turn conversations. Again, the promise is clear, but integration remains delicate: more reasoning can also mean more latency and more tokens consumed.
The agent guide provided by LG describes configurations for a generalist agent Hermes and a coding agent OpenCode. In the documented workflow, the external environment executes tool calls returned by the model, then sends their results back to the model for the continuation of the exchange. The model decides, the agent acts. This separation is classic but reminds us that the quality of an agentic system depends as much on orchestration as on the model itself.
Key points
- LG AI Research opens K-EXAONE 2.5 for free via API.
- Announcement dated November 9, 2026.
- The model claims 750 billion parameters.
- Supports Korean and English.
- K-EXAONE 2.5 targets code, research, and automation.
- LG compares its scores to DeepSeek-V4.1-Flash and GLM-5.3.
By the numbers
- 750 billion parameters — claimed model size, November 2026, LG AI Research.
- 524,288 tokens — input plus maximum generation limit, API documentation.
- 59.21 — SWE-Bench Pro score of K-EXAONE 2.5.
- 78.0 — BrowseComp score of K-EXAONE 2.5.
- 74.6 — AutomationBench open-set score of K-EXAONE 2.5.