LG AI Research has made its 750-billion-parameter K-EXAONE 2.5 model available to developers for free. In its November 9, 2026 announcement, the company is offering the model as a hosted service through the EXAONE API Platform. Developers, startups, companies, and research institutions that obtain an API key will be able to use the model for both product development and research work.
The new version is a technology validation model
LG describes K-EXAONE 2.5 as a technology validation model: it was produced by applying additional training on top of the version released in August 2026. The previous K-EXAONE 2.0 was an open-weight release; it was published under the Apache 2.0 license, used a Mixture-of-Experts architecture, and offered a 262,144-token context window with 37 billion active parameters. In the new version, the number of active parameters is listed as 31 billion, and the knowledge cutoff date is January 2026.
Strong results in agent tasks
According to LG’s own measurements, the model clearly pulls ahead of its predecessor in four agent-focused benchmarks. On SWE-Bench Pro, which measures software development and bug fixing, the score rises from 47.1 to 59.21. On BrowseComp, which measures finding information on the web, it goes from 50.87 to 78.0, and on AutomationBench, which measures open-ended tool use, it jumps from 16.9 to 74.6. On Claw-Eval, it increases from 62.11 to 77.29.
LG’s published comparison table also includes DeepSeek-V4.1-Flash and GLM-5.3. Based on those figures, K-EXAONE 2.5 is ahead of DeepSeek-V4.1-Flash on SWE-Bench Pro, BrowseComp, and Claw-Eval, but trails it on AutomationBench. Against GLM-5.3, it comes out ahead only on AutomationBench. The results are based on LG’s own statements and have not been verified by independent testing.
OpenAI-compatible interface and agent support
The model is text-only and supports Korean and English. It includes chat, tool calling, structured output, code generation, and flow features. The API uses the OpenAI-compatible POST /v1/chat/completions endpoint; the model ID is k-exaone_v2.5. If the total of input tokens and generated output exceeds 524,288, the request is rejected with a 400 error.
Reasoning intensity can be adjusted directly: reasoning_effort defaults to “max”, while preserve_thinking carries reasoning from previous turns into subsequent requests. LG’s agent guide includes ready-made configurations for the Hermes general agent and the OpenCode coding agent. In this setup, the model returns the tool calls, while executing those calls and managing the loop remain on the developer’s own infrastructure.
What free access provides
The free API makes it easy to test LG’s agent results in existing OpenAI-compatible toolchains. The real question is how well these internal benchmarks will translate to production software and workflows. Context management, tool execution, and reasoning depth will determine the model’s practical value.
Source: Unite.AI
Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.