{"slug": "lg-opens-k-exaone-2-5-api-free-claims-70-gain-2026", "title": "LG Opens K-EXAONE 2.5 API Free, Claims 70% Gain [2026]", "summary": "LG AI Research made the K-EXAONE 2.5 API free to all users on November 9, 2026, opening access through the EXAONE API Platform at docs.exaone.ai for developers, startups, enterprises and research institutions with no usage fee, announced at NVIDIA AI Day Seoul 2026 at COEX. LG AI Research said the technical-validation model, which received additional training on top of the version released in August, posted a performance gain of more than 70% over that earlier model across 24 evaluation metrics, and that it is pursuing technical cooperation with NVIDIA and considering NVIDIA Dynamo for inference. All performance figures come from LG's internal evaluation, and the company has not disclosed context window, token limits, rate limits, latency targets, serving hardware or an end date for free access.", "body_md": "LG AI Research made the **K-EXAONE 2.5 API** free for anyone on November 9, 2026. According to the [LG AI Research announcement](https://www.lgresearch.ai/news/view?seq=755), developers, startups, enterprises and research institutions can get an API key on the EXAONE API Platform at docs.exaone.ai and build services or run research with no usage fee. The company unveiled the move at NVIDIA AI Day Seoul 2026 at COEX in Seoul.\n\nThe release lands in a week when [open-weight models](https://shattered.io/open-weight-ai-models-side-by-side-the-spec-sheet/) from China keep setting the pace for cheap agentic AI. LG’s own tests, as Korean outlets relayed them, put the new model level with DeepSeek V4.1 Flash on several agent benchmarks. If you followed our earlier comparison of [Muse Spark Open, DeepSeek V4.1 Flash and Beam](https://shattered.io/muse-spark-open-vs-deepseek-v4-1-flash-vs-beam-1-0-1m-context-2026/), you already know how crowded that tier has become.\n\nThis analysis walks through what LG confirmed, what the benchmark table actually shows, where NVIDIA fits and where it does not, and what developers should test before they build on a free endpoint. One caution up front. LG describes K-EXAONE 2.5 as a technical-validation model, and every performance number so far comes from the company’s internal evaluation.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## What LG AI Research announced on November 9\n\nThe announcement has four parts. First, LG AI Research opened the K-EXAONE 2.5 API at no charge through its EXAONE API Platform. Second, it described the model as a technical-validation model that received additional training on top of the version released in August. Third, it reported a performance gain of more than 70% over that earlier model, measured across 24 evaluation metrics. Fourth, it said it is pursuing technical cooperation with NVIDIA and considering NVIDIA Dynamo for inference.\n\nThe [Korea Times](https://www.koreatimes.co.kr/business/companies/20261109/lg-ai-research-opens-k-exaone-25-api-free-of-charge) reported that access started the same Monday. It framed the model as part of the government’s sovereign AI foundation model project, which aims to build homegrown models that cut Korea’s reliance on foreign systems. Several Korean outlets carried near-identical details, which suggests they worked from the same LG release and briefing.\n\nThe timing matters too. NVIDIA AI Day Seoul 2026 gave LG a stage where hardware buyers, cloud teams and startups sit in the same room. Dropping a free API there signals that LG wants usage data and developer attention more than it wants API revenue this quarter.\n\n## How the K-EXAONE 2.5 API works and who can use it\n\nAccess runs through the EXAONE API Platform. Per LG’s release, a user obtains an API key and can then use the model for service development and research. The company lists developers, startups, enterprises and research institutions as eligible, and adds that anyone else can use it too.\n\n### Getting an API key\n\nThe announcement describes a simple path. You register on the platform, request a key and start calling the model. LG has not published, in the material reviewed here, any approval queue, usage tiers or commercial-use carve-outs. The free label in the release applies to usage fees, and the sources reviewed do not mention a paid plan.\n\n### What the announcement leaves open\n\nSeveral details that engineers care about are missing from the coverage. No outlet reported a context window, a token limit, rate limits, a latency target or the hardware serving the endpoint. Nobody reported whether the free access has an end date. Treat those as open questions until the platform documentation says otherwise, and plan your first tests around discovering them.\n\n## The 70% claim: what eight weeks and 24 metrics mean\n\nThe headline number is a performance gain of more than 70% in eight weeks. [Financial News](https://en.fnnews.com/news/202611091004295188) and Digital Today both reported that LG measured it by combining 24 evaluation metrics and comparing against the model released in August. The head of LG’s EXAONE Lab told reporters that an internal analysis confirmed the aggregate gain.\n\nThat wording deserves a careful read. An aggregate across 24 metrics is not a single accuracy score that went up 70%. LG did not publish, in the coverage reviewed, how it weighted the metrics, what the August baseline scored or which tasks drove the improvement. A model that starts weak on many tests can post a large relative gain while still trailing leaders in absolute terms.\n\nOne small discrepancy shows up in the coverage. The Korea Times quoted the comparison point as the model released in July, while LG’s own release and the other outlets say August. We use August here because it comes from the company’s release and most of the reporting.\n\nDigital Today explained that the improvement came from post-training, the extra training stage after pre-training that sharpens reasoning, problem-solving and instruction following. LG refined its training methodology rather than building a new base model. That distinction explains how a gain of this size could arrive in eight weeks.\n\n## K-EXAONE 2.5 benchmarks vs DeepSeek V4.1 Flash and GLM 5.3\n\nLG tested the model on four agentic AI benchmarks using public datasets. It compared the results with two Chinese [open-weight](https://shattered.io/deepseek-vs-qwen-vs-mistral-open-weight-models-2026/) models, DeepSeek V4.1 Flash and GLM 5.3 (max). The table below combines the scores reported by [Digital Today](https://www.digitaltoday.co.kr/en/view/122771/lg-ai-research-opens-k-exaone-2-5-api-free-says-performance-up-70-percent) with the benchmark descriptions in LG’s release. All of them are company-run results.\n\n| Benchmark | What it measures | K-EXAONE 2.5 | DeepSeek V4.1 Flash | GLM 5.3 (max) | \n|---|---|---|---|---|\n| AutomationBench | Task automation with SaaS tools | 74.6 | 74.9 | 71.6 | \n| Claw-Eval | Agent problem-solving in workplace settings | 77.3 | 75.8 | 77.6 | \n| BrowseComp | Web browsing and information finding | 78.0 | 75.8 | 79.0 | \n| SWE-Bench Pro | Real-world software modification | 59.2 | 58.4 | 64.0 | \n\nTreat the table as a vendor scorecard. LG ran the evaluations itself, and the coverage does not mention independent replication, evaluation settings or harness versions. Agentic benchmarks swing noticeably with prompt format and tool configuration, so small gaps carry little weight.\n\n## Reading the scores: where K-EXAONE 2.5 wins and loses\n\nAgainst DeepSeek V4.1 Flash, the reported numbers favor LG on three of four tests. The lead on Claw-Eval is 1.5 points, on BrowseComp 2.2 points and on SWE-Bench Pro 0.8 points. On AutomationBench, LG trails by 0.3 points, which Financial News called comparable. Those gaps sit within the range where a different harness could flip the order.\n\nAgainst GLM 5.3 (max), the picture reverses. K-EXAONE 2.5 wins only AutomationBench, by 3.0 points. GLM leads on Claw-Eval by 0.3, on BrowseComp by 1.0 and on SWE-Bench Pro by 4.8 points. The coding gap is the largest in the table, and coding is where many developers will test first.\n\nA simple unweighted average of the four reported scores, our own arithmetic and not an official index, puts K-EXAONE 2.5 at 72.8, DeepSeek V4.1 Flash at 71.2 and GLM 5.3 (max) at 73.1. The benchmarks measure different skills, so treat that as a rough guide. LG’s model lands between the two rivals, close to both, which fits the company’s stated claim of being competitive rather than leading.\n\nFor context on how the DeepSeek line performs in offensive-security tasks, see our write-up on [the open model that scored 72.8% on the Vals hacking test](https://shattered.io/vals-cyberoffense-deepseek-v4-1-flash-72-8-percent-vs-gpt-6-sol-2026/).\n\n## NVIDIA AI Day Seoul and the Dynamo question\n\nLG said at the event that it is pursuing technical cooperation with NVIDIA. It also said it is considering NVIDIA Dynamo to raise inference efficiency and service scalability. The sources reviewed do not confirm a signed partnership, a joint product or a Dynamo deployment. Headlines that say LG partnered with NVIDIA run ahead of the evidence.\n\nDigital Today described Dynamo as NVIDIA’s [open-source](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/) inference optimization software that speeds up model serving by allocating GPU resources efficiently. It also reported that NVIDIA had earlier released a Dynamo usage guide tailored to K-EXAONE. That detail suggests the two sides have already done some technical groundwork, though it does not prove a commercial arrangement.\n\nWhy would Dynamo matter for a free API? Inference cost is the line item that decides whether a zero-fee endpoint stays viable. Better GPU scheduling lets one cluster serve more requests, which lowers the cost of every free call. If LG adopts it, the benefit shows up in latency and availability rather than in benchmark scores.\n\nFor readers tracking the serving side of this story, our coverage of [Intel’s day-0 support for Muse Spark Open on four GPUs](https://shattered.io/intel-day-0-support-meta-muse-spark-open-816b-4-gpus-160gb-2026/) shows how much hardware planning goes into running a large open model.\n\n## Why give the API away? The economics of a zero-fee model\n\nLG has not explained its reasoning beyond the stated goals. Reports say it wants to verify the model’s performance and usability in real service environments and widen the use of its proprietary model inside Korea’s AI ecosystem. The Korea Times also tied the release to the sovereign AI project’s goal of reducing dependence on foreign models.\n\nRead between the lines and a free API works as a large-scale test. Thousands of developers hitting an endpoint with real prompts will expose failure modes that internal benchmarks miss. LG says it plans to feed the training techniques it validated here into its third-phase model, including work on data quality, post-training and reinforcement learning. Usage data from the API could help guide that work, although LG has not said it will use the data that way.\n\nThere is a competitive logic as well. Chinese open-weight models gained ground partly because developers could try them cheaply. A free tier from a Korean lab lowers the switching cost for teams that prefer a domestic provider, whether for language fit, data-residency preferences or procurement rules. Whether the model is good enough to win those teams is the real test.\n\n## Korea’s sovereign AI race: from K-EXAONE 2.0 to a third model\n\nK-EXAONE 2.5 sits inside a government-backed competition. According to [Maeil Business Newspaper](https://www.mk.co.kr/en/business/12202406), the K-LLM contest has narrowed to three contenders, LG AI Research, SK Telecom and Upstage, with two companies to be selected next year. That structure explains the urgency behind an eight-week improvement cycle and a public API launch.\n\n| Item | What reports say | Source type | \n|---|---|---|\n| K-EXAONE 2.0 | Second-phase model of the sovereign AI foundation model project, released in August | Korea Times, Digital Today, LG release | \n| K-EXAONE 2.5 | Technical-validation model with extra post-training, API opened free on November 9 | LG release and Korean outlets | \n| Third-phase model | LG aims to complete a frontier-level EXAONE model by January next year | Financial News, Digital Today | \n| K-LLM contenders | LG AI Research, SK Telecom and Upstage, with two to be selected next year | Maeil Business Newspaper | \n\nLG’s co-head of research said the lab is concentrating on the third model and hopes the API lets more developers and companies see the improved performance firsthand. He repeated the company’s goal of finishing a frontier-level EXAONE model by next January. That is a stated target and not a delivered product, so the usual caution about timelines applies.\n\n## What the release means for developers and startups\n\nFor a small team, a free agent-capable model changes prototyping math. You can wire an agent loop, a browsing tool or a code-assist feature to the endpoint without watching a token bill. That works best for internal tools, research projects and early product experiments where cost uncertainty slows decisions.\n\nProduction use is a different question. A free endpoint from a validation model carries service-level uncertainty, since LG has not published uptime commitments, rate limits or a deprecation policy in the material reviewed. Teams building anything customer-facing should keep a fallback model wired in and avoid hard-coding assumptions about availability. Our look at [Forrester’s three limits on agent spending](https://shattered.io/ai-bill-shock-forrester-three-agent-limits-openai-invoice-2026/) covers the cost discipline that still applies even when the sticker price is zero.\n\nResearchers gain a different kind of value. A well-documented free API from a national lab makes it easier to reproduce studies and compare behavior across models. Korean-language and Korea-specific tasks may also benefit, though LG’s coverage did not publish language benchmarks, so that remains a hypothesis to test.\n\n## K-EXAONE 2.5 API vs other ways to use open-weight models\n\nDevelopers now have three broad routes to an open-weight agent model. They can call a hosted API, rent a managed cloud plan or self-host on their own GPUs. LG’s offer adds a free hosted option with a national-lab brand attached. It does not change the tradeoffs among the three.\n\n- **Hosted free API:** zero fees and fast start, but you inherit LG’s limits and uptime, and you send prompts to a third party.\n- **Managed cloud plans:** predictable service terms and support, with a monthly or per-token cost.\n- **Self-hosting:** full control over data and latency, with hardware, ops and electricity on your side.\n\nIf you are weighing the self-hosting route, our comparison of [Ollama Cloud pricing against self-hosted vLLM](https://shattered.io/ollama-cloud-pricing-vs-self-hosted-vllm-20-pro-plan-2026/) lays out the cost break-even logic. A free endpoint resets that math for experiments, but not for workloads that need guaranteed capacity or strict data handling.\n\nOne more factor is model weights. The sources reviewed describe the API release only and do not say LG is publishing weights for K-EXAONE 2.5. LG’s comparison group is open-weight models, but the announcement as reported concerns API access. Do not assume you can download and self-host the model.\n\n## Confirmed, reported and unconfirmed\n\nFast-moving AI launches attract loose wording, so a plain sorting helps. The table separates what LG said directly from what depends on LG’s own tests and from what nobody has shown.\n\n| Claim | Status | Basis | \n|---|---|---|\n| API is free via the EXAONE API Platform | Confirmed by the company | LG AI Research release | \n| Announced November 9 at NVIDIA AI Day Seoul 2026, COEX | Confirmed | LG release and multiple outlets | \n| Gain of more than 70% across 24 metrics since August | Company-reported, internal | Financial News, Digital Today | \n| Competitive with DeepSeek V4.1 Flash and GLM 5.3 (max) | Company-reported, internal | LG release, Korea Times | \n| NVIDIA technical cooperation | Pursued, per LG | Financial News, Digital Today | \n| NVIDIA Dynamo in use | Under consideration only | Same outlets | \n| Context window, rate limits, hardware, pricing after launch | Not disclosed in coverage reviewed | None | \n| NVIDIA as a formal partner | Unconfirmed | None | \n\n## Risks and open questions\n\nFour questions stand out. First, how long does free access last? A launch promotion and a permanent policy look the same on day one. Second, what happens to the data you send? The coverage reviewed does not describe data retention or training use, so read the platform terms before sending anything sensitive.\n\nThird, will independent testers confirm LG’s numbers? Self-reported agentic scores often shrink when outside labs rerun them. Fourth, can LG serve demand? A free API with a strong brand can draw heavy traffic fast, and without published rate limits it is hard to predict how the platform behaves under load.\n\nThere is also a geopolitical thread. LG chose DeepSeek and GLM as its yardsticks, which places the model in a global race rather than a purely domestic one. Our earlier analysis of [how a Democratic win could open an AI window for China](https://shattered.io/democrat-win-china-ai-window-to-catch-up-natixis-2026/) shows how closely markets watch that race. A Korean lab matching Chinese open models on selected tests will draw attention, but one internal scorecard does not settle the contest.\n\n## Five predictions for K-EXAONE 2.5 and the EXAONE platform\n\nThese are our analytical judgments, not reported facts, and each one could prove wrong.\n\n1. **Independent benchmark runs will land within a few points of LG’s numbers, and the coding gap to GLM will stay visible.** SWE-Bench Pro showed the widest gap in LG’s own table, so outside runs are unlikely to close it.\n2. **LG will publish usage terms or limits within weeks.** Free platforms without stated limits tend to add them once real traffic arrives.\n3. **NVIDIA cooperation will stay at the technical level through year-end.** The reported language is about pursuing cooperation and considering Dynamo, which points to engineering work rather than a commercial announcement.\n4. **The third-phase model will become the real headline.** LG set a January target, and the K-LLM selection next year will push the company to show results on a public schedule.\n5. **Rivals in the K-LLM contest will answer with their own developer access programs.** SK Telecom and Upstage compete for the same two slots, and developer mindshare is cheap to chase.\n\n## A short checklist for testing the model this week\n\nIf you plan to try the API, test it the way you would test any new vendor. Start by confirming the platform terms, then run your own tasks instead of trusting public leaderboards. A small harness that replays 50 to 100 of your real prompts will tell you more than a headline score.\n\n- Read the platform terms for data retention, commercial use and rate limits before the first call.\n- Replay real prompts from your product and score outputs against your current model.\n- Measure latency at different hours, since a free endpoint may vary under load.\n- Test tool-calling and multi-step agent flows, because that is where LG claims strength.\n- Keep a paid fallback model configured so an outage does not break your product.\n\nFor broader coverage of model launches and benchmarks, browse our [AI and machine learning section](https://shattered.io/ai-machine-learning/).\n\n## The takeaway: a free API is a bet on adoption, not proof of quality\n\nLG AI Research gave away its K-EXAONE 2.5 API, claimed a 70% gain in eight weeks and put the model in the same conversation as DeepSeek V4.1 Flash and GLM 5.3. Those are meaningful moves for a Korean lab competing in a government-run selection. The evidence behind them is company-run, and the NVIDIA story is cooperation under discussion rather than a finished deal.\n\nDevelopers should treat the free endpoint as a low-risk way to test a new agent model and a poor basis for production commitments until LG publishes limits and terms. The next signals to watch are independent benchmark runs, platform documentation updates and progress on the third-phase model that LG wants to finish by January.\n\n## FAQ: K-EXAONE 2.5 API\n\n### Is the K-EXAONE 2.5 API really free?\n\nAccording to LG AI Research, yes. Users can obtain an API key on the EXAONE API Platform and use the model for service development and research at no cost. The sources reviewed do not describe paid tiers or an end date.\n\n### Where do I get access?\n\nThrough the EXAONE API Platform at docs.exaone.ai. LG says developers, startups, companies, research institutions and other users can request a key.\n\n### What does the 70% improvement mean?\n\nLG reported that performance combined across 24 evaluation metrics rose by more than 70% in eight weeks, compared with the model released in August. It is an aggregate from LG’s internal analysis, and the sources reviewed do not publish the baseline or weighting.\n\n### How does K-EXAONE 2.5 compare with DeepSeek V4.1 Flash?\n\nIn LG’s own tests, it scored higher on Claw-Eval, BrowseComp and SWE-Bench Pro and slightly lower on AutomationBench, 74.6 against 74.9. These are company-run results on public datasets and lack independent confirmation.\n\n### Did NVIDIA become an official partner of LG AI Research?\n\nThat is not confirmed. LG said it is pursuing technical cooperation with NVIDIA and considering NVIDIA Dynamo for inference. No formal partnership or Dynamo deployment has been reported.\n\n### Can I download K-EXAONE 2.5 and run it myself?\n\nThe coverage reviewed describes API access only. It does not say LG is releasing model weights, so do not plan a self-hosted deployment on the strength of this announcement.\n\n### What is K-EXAONE 2.5 in LG’s model lineup?\n\nLG calls it a technical-validation model built by adding training to the August release, which is the second-phase model of Korea’s sovereign AI foundation model project. LG plans to apply the validated techniques to its third-phase model and aims to finish a frontier-level version by January next year.", "url": "https://wpnews.pro/news/lg-opens-k-exaone-2-5-api-free-claims-70-gain-2026", "canonical_source": "https://shattered.io/lg-ai-research-k-exaone-2-5-api-free-70-percent-nvidia-2026/", "published_at": "2026-11-09 09:56:30+00:00", "updated_at": "2026-11-09 10:22:55.839573+00:00", "lang": "en", "topics": ["large-language-models", "ai-products", "ai-tools", "ai-infrastructure", "ai-startups"], "entities": ["LG AI Research", "K-EXAONE 2.5", "EXAONE API Platform", "NVIDIA", "NVIDIA Dynamo", "NVIDIA AI Day Seoul 2026", "DeepSeek V4.1 Flash", "Korea Times"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/lg-opens-k-exaone-2-5-api-free-claims-70-gain-2026", "markdown": "https://wpnews.pro/news/lg-opens-k-exaone-2-5-api-free-claims-70-gain-2026.md", "text": "https://wpnews.pro/news/lg-opens-k-exaone-2-5-api-free-claims-70-gain-2026.txt", "jsonld": "https://wpnews.pro/news/lg-opens-k-exaone-2-5-api-free-claims-70-gain-2026.jsonld"}}