K-EXAONE 2.5 explained: free API, 750B, vendor-reported gains

By LLMTest Lab · Nov 9, 2026 · 4 min read releasellmopen-weight
On this page

On this page

  1. What is new
  2. What LG reports
  3. Who should care
  4. How to try it

Not tested by us yet. We have not run K-EXAONE 2.5 through our harness: LG's API needs a key from its own platform, and the model is not listed on OpenRouter, which is where our test runs go. Everything below comes from LG's EXAONE site and API documentation, plus coverage by Unite.AI and the Korea Times. We will add our own numbers once we have a key and a stable endpoint.

K-EXAONE 2.5 is a text-only model from LG AI Research with 750 billion total parameters, opened for free API use on November 9, 2026. LG says it scores on par with DeepSeek-V4.1-Flash and GLM-5.3 on four agentic benchmarks, all measured by LG itself. If you build agents and want a no-cost model to try, it is worth a trial. If you need a stable price or open weights, there is nothing to plan around yet.

K-EXAONE 2.5
Maker LG AI Research
Announced 2026-11-09
Parameters 750B total, 31B active (per model docs, via Unite.AI)
Modality Text only, Korean and English
Knowledge cutoff January 2026 (per Unite.AI)
API OpenAI-compatible POST /v1/chat/completions, model id k-exaone_v2.5
Token limit Input plus max_tokens up to 524,288
Price Free, no end date published
Weights Not announced

What is new

LG calls K-EXAONE 2.5 a "technology-validation model", produced by further training of K-EXAONE 2.0. According to the Korea Times, LG's internal analysis shows a gain of more than 70 percent across 24 benchmarks compared with the earlier model, achieved in about eight weeks. The two outlets disagree on whether 2.0 shipped in July or August, so treat the exact gap as approximate.

K-EXAONE 2.0 is the open-weight predecessor under Apache 2.0, with 37B active parameters and a 262,144-token context, according to Unite.AI. K-EXAONE 2.5 doubles the usable token limit on the API and has fewer active parameters, which should matter for latency, though LG has not published speed figures.

The API exposes reasoning_effort, which defaults to max, and preserve_thinking to carry earlier reasoning into multi-turn requests. LG warns that higher effort raises response time and token use. Since the model is free, the cost shows up as latency rather than money.

What LG reports

These are LG's own measurements on public benchmarks, as listed by Unite.AI. Nobody independent has reproduced them yet.

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

Two readings of that table. First, the jump over 2.0 is large, especially on AutomationBench, where 74.6 against 16.9 is the source of the "4.4x" headline that circulates in the press. That ratio is a publication's arithmetic on LG's numbers, not a separate result. Second, against the two named rivals the picture is mixed: 2.5 beats DeepSeek-V4.1-Flash on three of four tests, loses to GLM-5.3 on three of four, and every gap is a point or two on a single vendor-run test. "Roughly level" is a fairer summary than "ahead".

Who should care

Teams building agents on a budget have the clearest reason. A free, OpenAI-compatible endpoint means you can swap the base URL in an existing client and run your own evaluation at no cost. LG provides harness configurations for the Hermes general agent and the OpenCode coding agent, so you can try tool-calling workflows without writing glue code.

Korean-language products are the second group, since LG positions the model around Korean understanding. We have no Korean test of our own, so the claim stands unverified.

Everyone else should wait. LG has not said how long access stays free, what it will charge afterwards, or whether the 2.5 weights will be released. A free tier with no end date can still disappear, and a production pipeline built on it has no fallback price. LG has said it plans to apply what it learned to a third-round model aimed at frontier-level performance by January, so this release may be short-lived.

How to try it

Request a key on the EXAONE API platform, set the model to k-exaone_v2.5, and run the same prompts you use to pick between your current models. Our guide on how to choose an LLM covers building that comparison set. If privacy or self-hosting matters more than a free endpoint, the open-source LLMs for on-prem use roundup is the better starting point, and for another recent open-weight release from a non-US lab, see our Arete-27B explainer.

We will update this post when we can test the API ourselves.

Related articles

Arete-27B explained: Nous Research's re-aligned Qwen3.8 27B
Arete-27B is Nous Research's Apache-2.0 fine-tune of Qwen3.8-27B, built to cut state-propaganda answers. Specs, vendor-reported evals, caveats, who should care.
Best open-source LLMs for on-prem in 2026: what actually runs where
Six open-source LLMs ranked for on-prem deployment in 2026: hardware minimums, real license terms, and the performance tier you get at each level.