Innovation & Tech

LG AI Research Opens Free API for 750-Billion-Parameter K-EXAONE 2.5

LG AI Research opened free hosted API access to K-EXAONE 2.5 on November 9, 2026. Internal benchmarks show broad gains, but no production results are disclosed.

By Olivia Hart

4 min read

LG AI Research Opens Free API Access to Its K-EXAONE 2.5 Model
LG AI Research Opens Free API Access to Its K-EXAONE 2.5 ModelAkela999 / pixabay

What's News

  • LG AI Research announced free API access to the 750-billion-parameter K-EXAONE 2.5 on November 9, 2026.
  • LG's documentation lists 31 billion active parameters, a January 2026 knowledge cutoff, and Korean and English support.
  • K-EXAONE 2.5 scored 59.21 on SWE-Bench Pro, 78.0 on BrowseComp, 74.6 on AutomationBench open-set and 77.29 on Claw-Eval.
  • The API caps input plus max_tokens at 524,288 tokens; larger requests return a 400 error.
  • LG reported gains over K-EXAONE 2.0 on all four listed benchmarks using internal open-dataset tests.

By Olivia Hart

What does LG's free API include?

On November 9, 2026, LG AI Research announced free API access to the 750-billion-parameter K-EXAONE 2.5 model through the EXAONE API Platform. Developers, startups, companies and research institutions can obtain an API key for hosted model use in service development and research. LG did not state how long free access will last or what usage limits apply.

LG described K-EXAONE 2.5: "a technology-validation model produced by applying additional training to the version released in August 2026." The earlier K-EXAONE 2.0 model card documents an open-weight release under Apache 2.0, with a Mixture-of-Experts architecture, 37 billion active parameters and a 262,144-token context length.

LG does not say whether K-EXAONE 2.5 itself carries an open-weight license. Takeaway: Free hosted access lowers the cost of testing, but it does not establish that K-EXAONE 2.5 is open-weight.

What do the vendor-reported benchmarks show?

LG AI Research said it measured K-EXAONE 2.5 internally on open datasets and found the model competitive with frontier-class global open-weight models. Its documentation compares K-EXAONE 2.5 with K-EXAONE 2.0, DeepSeek-V4.1-Flash and GLM-5.3:

  • SWE-Bench Pro, agentic coding: K-EXAONE 2.5 scored 59.21; K-EXAONE 2.0, 47.1; DeepSeek-V4.1-Flash, 58.41; GLM-5.3, 64.03.
  • BrowseComp, agentic search: K-EXAONE 2.5 scored 78.0; K-EXAONE 2.0, 50.87; DeepSeek-V4.1-Flash, 75.8; GLM-5.3, 79.0.
  • AutomationBench open-set, agentic tool use: K-EXAONE 2.5 scored 74.6; K-EXAONE 2.0, 16.9; DeepSeek-V4.1-Flash, 74.9; GLM-5.3, 71.6.
  • Claw-Eval, agentic tool use: K-EXAONE 2.5 scored 77.29; K-EXAONE 2.0, 62.11; DeepSeek-V4.1-Flash, 75.78; GLM-5.3, 77.64.

LG said the benchmarks cover software development and modification, finding information on the web, automating work with SaaS tools, and agent problem-solving in work environments. On LG's figures, K-EXAONE 2.5 improved on K-EXAONE 2.0 across all four tests.

The model scored above DeepSeek-V4.1-Flash on SWE-Bench Pro, BrowseComp and Claw-Eval, but below it on AutomationBench. Against GLM-5.3, it scored higher only on AutomationBench. LG named code generation, debugging, refactoring and test-code writing as applications, making software development central alongside agentic search and workplace automation.

These comparisons do not show a universal lead over external models. Takeaway: LG's improvement over its predecessor is broad, but its edge over rivals depends on the task.

What can developers run through the API?

K-EXAONE 2.5 is text-only and supports Korean and English. LG's documentation lists 31 billion active parameters for the 750-billion-parameter model, against 37 billion for K-EXAONE 2.0, along with a January 2026 knowledge cutoff, chat, tool calling, structured outputs, code generation and streaming.

The endpoint reference specifies an OpenAI-compatible POST /v1/chat/completions interface using the model identifier k-exaone_v2.5. Input tokens plus the max_tokens generation allowance, including reasoning, cannot exceed 524,288 tokens. Larger requests return a 400 error.

Responses use standard OpenAI-format JSON, or server-sent-event chunks when streaming is enabled. Reasoning controls are exposed directly: reasoning_effort defaults to max, while preserve_thinking can carry earlier assistant reasoning into multi-turn requests.

LG's agent guide says: "higher reasoning intensity can increase response time and token usage." It treats current-response effort separately from passing earlier reasoning and tool-call history. Takeaway: The API supports common development patterns, but maximum reasoning can consume more time and tokens.

How does agent integration work?

LG's agent guide includes configurations for the Hermes general agent and the OpenCode coding agent. In the documented workflow, an external harness executes tool calls returned by the model and passes their results back while handling text, reasoning and tool-call deltas across subsequent turns.

Developers therefore receive the model's reasoning and tool-call decisions through the API, but the surrounding agent remains responsible for executing actions and managing the loop. LG says free access lowers the cost of testing agentic-coding results inside existing OpenAI-compatible tooling.

The announcement names no external customer or production deployment with measured results. Takeaway: The model can direct an agent workflow, but outside software must execute actions and keep each loop moving.

What remains unproven?

LG's benchmark results come from internal measurements on open datasets, not independent production validation. The announcement cites no external replication, production latency or total cost per completed software task.

Free access lowers the barrier to testing LG's claims inside tools developers already use. LG framed the unresolved issue as whether those internally measured gains transfer into production software and business workflows, where reasoning depth, tool execution and context management will shape practical value.

Those variables will matter more as developers move from benchmark tables into live systems. Takeaway: Free access can broaden the test pool, but measured production performance will determine whether LG's benchmark gains become durable commercial value.

Source: unite.ai

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,205 articles

Related articles

« Previous articleNext article »