LG opens free API access to its 750B-parameter K-EXAONE 2.5

By Elena Vasquez

On November 9, 2026, LG AI Research announced free API access to K-EXAONE 2.5, a 750-billion-parameter model available through the EXAONE API Platform.

LG says developers, startups, companies and research institutions can obtain an API key for service development and research.

LG's announcement says free hosted access lowers the cost of testing its agentic-coding results inside existing OpenAI-compatible tooling.

What changed on November 9?

LG describes K-EXAONE 2.5 as a “technology-validation model” produced by applying additional training to the version released in August 2026.

Its predecessor, K-EXAONE 2.0, has an open-weight release under the Apache 2.0 license, according to the earlier model card.

That card also records a Mixture-of-Experts architecture, 37 billion active parameters and a 262,144-token context length.

LG's current documentation lists 31 billion active parameters for K-EXAONE 2.5 and a January 2026 knowledge cutoff.

LG separately describes the new release as a 750-billion-parameter model.

LG's description makes the lineage incremental: the August release received additional training before the November API offering.

The two active-parameter figures belong to different releases, while the model card and API documentation also use different context and request limits.

That distinction matters for developers comparing the open-weight 2.0 release with the hosted 2.5 service, because specifications and model generations do not carry over unchanged.

What do LG's agent benchmarks show?

LG says it measured K-EXAONE 2.5 internally on open datasets.

The company describes the results as competitive with frontier-class global open-weight models.

LG's benchmark table reports:

• SWE-Bench Pro, agentic coding: K-EXAONE 2.5 scored 59.21, compared with 47.1 for K-EXAONE 2.0, 58.41 for DeepSeek-V4.1-Flash and 64.03 for GLM-5.3. The test addresses software development and modification.
• BrowseComp, agentic search: K-EXAONE 2.5 scored 78.0, versus 50.87 for K-EXAONE 2.0, 75.8 for DeepSeek-V4.1-Flash and 79.0 for GLM-5.3. The benchmark measures finding needed information on the web.
• AutomationBench open-set, agentic tool use: K-EXAONE 2.5 scored 74.6, compared with 16.9 for K-EXAONE 2.0, 74.9 for DeepSeek-V4.1-Flash and 71.6 for GLM-5.3. The test covers automating work with SaaS tools.
• Claw-Eval, agentic tool use: K-EXAONE 2.5 scored 77.29, versus 62.11 for K-EXAONE 2.0, 75.78 for DeepSeek-V4.1-Flash and 77.64 for GLM-5.3. The benchmark addresses agent problem-solving in work environments.

On LG's reported figures, K-EXAONE 2.5 improves on K-EXAONE 2.0 across all four benchmarks.

It scores above DeepSeek-V4.1-Flash on SWE-Bench Pro, BrowseComp and Claw-Eval, but below it on AutomationBench open-set.

Against GLM-5.3, K-EXAONE 2.5 scores higher only on AutomationBench open-set.

The ordering remains consistent across three tests: K-EXAONE 2.5 sits between its predecessor and GLM-5.3 on SWE-Bench Pro, BrowseComp and Claw-Eval.

The internal table therefore presents a clear gain over LG's own model across all four tasks, but no single competitor leads every benchmark in the table.

What can developers build with it?

LG names code generation, debugging, refactoring and test-code writing as applications.

Software development is therefore a central use case, alongside agentic search and workplace automation, according to LG.

K-EXAONE 2.5 is text-only and supports Korean and English, LG's documentation says.

LG lists chat, tool calling, structured outputs, code generation and streaming among the model's features.

For developers, that combination offers both a direct chat endpoint and the building blocks for agents that call tools, return structured data and stream partial results.

The feature set supports both interactive assistance and agentic workflows, with software development remaining the anchor use case in LG's release.

How does the API handle requests?

LG's endpoint reference specifies an OpenAI-compatible POST /v1/chat/completions interface and the model identifier k-exaone_v2.5.

Input tokens plus the max_tokens generation allowance, which includes reasoning, cannot exceed 524,288 tokens.

Requests above that combined limit return a 400 error.

Responses use standard OpenAI-format JSON, or server-sent-event chunks when streaming is enabled.

The documented ceiling counts both input and maximum generated output, including reasoning, rather than acting as a context-length label alone for this hosted release.

Existing OpenAI-compatible tooling can send requests to the documented endpoint and consume either JSON responses or streaming chunks.

How much control do developers have over reasoning?

LG exposes reasoning controls directly through the API.

The reasoning_effort setting 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.”

The guide treats current-response effort separately from passing earlier reasoning and tool-call history.

Developers can therefore tune the current answer's effort while deciding separately whether the next turn carries forward earlier reasoning and tool calls from the conversation.

That separation gives developers control over both computation and retained context, although LG notes that the more intensive setting comes with higher latency and token consumption.

Who executes the agent's actions?

LG's 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 K-EXAONE 2.5 and passes their results back to the model.

The harness also handles text, reasoning and tool-call deltas across subsequent turns.

Developers receive reasoning and tool-call decisions through the API, while the surrounding agent remains responsible for executing actions and managing the loop.

The division matters because model quality alone does not complete a workflow: the harness must execute each action, return its result and preserve the conversation's state.

In LG's setup, the model chooses and structures actions, while the surrounding harness manages the operational work those actions require.

What will free access prove?

LG's benchmark results come from internal measurements on open datasets rather than a named production deployment.

According to the announcement, free access lowers the cost of testing LG's agentic-coding gains inside existing OpenAI-compatible systems.

The practical test is whether LG's internally measured gains carry into production software and business workflows, where reasoning depth, tool execution and context management shape practical value.

Developers, startups, companies and research institutions can now submit those tests without first arranging a separate hosted deployment of K-EXAONE 2.5.

The free API now puts that test in developers' hands; wider use will show whether LG's reported gains translate into reliable production results.