LG AI Research opened free API access to K-EXAONE 2.5 on November 9, 2026, offering hosted use of a 750-billion-parameter model through the EXAONE API Platform. Developers receive an API key for service development and research, with requests handled through an OpenAI-compatible interface. The release centers on software development, agentic search and workplace automation. This matters because teams can test LG-reported gains in coding and agent tasks without changing existing tooling.
Free hosted access through EXAONE API Platform
Access is available to developers, startups, companies and research institutions that obtain an API key for the EXAONE API Platform. LG describes K-EXAONE 2.5 as a technology-validation model produced by additional training of the version released in August 2026. The model is text-only and supports Korean and English, with documentation listing 31 billion active parameters. Its knowledge cutoff is January 2026. Listed features include chat, tool calling, structured outputs, code generation and streaming.
The endpoint uses POST /v1/chat/completions with the model identifier k-exaone_v2.5 and returns standard OpenAI-format JSON, or server-sent-event chunks when streaming is enabled. Input tokens plus the max_tokens allowance, which includes reasoning, cannot exceed 524,288 tokens, and larger requests return a 400 error. Reasoning controls are exposed directly, with reasoning_effort defaulting to max. The preserve_thinking option can carry earlier assistant reasoning into multi-turn requests. LG notes that higher reasoning intensity can increase response time and token usage.
LG said internal measurement on open datasets showed results competitive with frontier-class global open-weight models. The documented comparison covers software development and modification, finding information on the web, automating work with SaaS tools, and agent problem-solving in work environments. On the reported figures, K-EXAONE 2.5 improves on its predecessor across all four benchmarks. It scores above DeepSeek-V4.1-Flash on SWE-Bench Pro, BrowseComp and Claw-Eval, but below it on AutomationBench. Against GLM-5.3, it scores above that model only on AutomationBench.
What testing means for automation teams
For companies already using OpenAI-compatible pipelines, the practical effect is lower cost of experimentation with an agent-oriented coding model. Named applications include code generation, debugging, refactoring and test-code writing, alongside agentic search and workplace automation. A small team can run trials inside current harnesses without procuring weights or separate infrastructure. A larger organization can compare the model side by side with incumbents on its own repositories, SaaS workflows and support knowledge.
The surrounding agent loop remains the responsibility of the customer system. In the documented workflow, the external harness executes tool calls returned by the model and passes results back, while handling text, reasoning and tool-call deltas across turns. The guide includes configurations for the Hermes general agent and OpenCode coding agent. Limits to verify include the 524,288-token combined input and generation budget, text-only input, and Korean and English support. The benchmark figures are internally measured, so they do not by themselves confirm production behavior.
A useful marker will be independent deployment reports showing the same pattern on SWE-Bench Pro, BrowseComp, Claw-Eval and AutomationBench-type work inside real codebases and SaaS environments. If teams reproduce faster debugging cycles and stable tool calling under the stated token and reasoning settings, the free API will have converted a validation release into a procurement option. If not, it will remain a low-cost benchmark reference.
