LG AI Research released EXAONE 4.0 on 15 July 2025 in 32-billion-parameter and 1.2-billion-parameter versions and began a commercial API service with FriendliAI. On the same day, it gathered 22 domestic partners to discuss the surrounding ecosystem. The launch combined a new model family with a way for developers to use it without operating high-performance GPUs.
The important decision was to give EXAONE more than one route into work. The downloadable models supported research and education; the hosted API supplied execution; a subsequent on-premise package addressed companies wanting systems inside their own networks. That combination followed an employee service already in operation. LG was building on practical internal distribution while asking outside partners to carry the model further.
FriendliAI's API removed the initial GPU-installation task from developers trying EXAONE. The later on-premise package took the model in the other direction, into a company's own network. LG was building several delivery formats around the same research programme, drawing on its employee service and partnerships in model serving and inference hardware.
A Product Launch With Partners Already in the Room
The launch came four months after EXAONE Deep, LG's reasoning-focused release. EXAONE Lab head Jinsik Lee said the team would keep developing the family as a Korean frontier model. The new version brought ordinary language responses and deliberate problem-solving into one model family. Instead of presenting those as separate destinations for a customer, LG put the choice inside the same generation of technology.
LG introduced the two sizes and FriendliAI's hosted service on the same day. Its partners' meeting brought twenty-two domestic organisations together to discuss how to use and support EXAONE. Jinsik Lee's new model was being launched with a distribution partner and a wider group of potential collaborators. A larger public presentation was scheduled for 22 July at LG Sciencepark in Magok.
Two Modes and Two Sizes Changed the Starting Point
The original technical report, submitted on 15 July, connects the new model to two earlier strands of work: EXAONE 3.5's instruction-following and EXAONE Deep's reasoning. It adds tool use and Spanish to Korean and English. These were specific development choices aimed at applications, not simply an increase in parameter count.
The report describes a 32B model for higher-performance work and a 1.2B model for devices. Its configuration table gives maximum sequence lengths of 131,072 and 65,536 tokens respectively. The larger architecture combined local and global attention to handle longer inputs. Developers could select an ordinary response or a reasoning mode, while the smaller model offered another starting point for phones, appliances, vehicle electronics and robots.
Parameter count of the larger EXAONE 4.0 model released in July 2025.
Source 2The Employee Service Came Before the External Offer
The commercial launch did not begin with an empty internal user base. On 9 December 2024, LG had opened ChatEXAONE to employees through a dedicated website alongside EXAONE 3.5. That release offered 2.4B, 7.8B and 32B models. The employee service combined web information and uploaded documents with summarisation, translation, report-writing, data analysis and coding.
It supplied recommended prompts across fourteen job categories and 133 tasks, with employee feedback intended to expand those classifications. Its Deep function broke compound questions into stages; Dive let users choose categories of sources, including overseas sites, academic material and YouTube. The service put EXAONE inside a working interface with document access, security and task-specific prompts. July's model arrived after that internal introduction.
The Following Week Made the Enterprise Package Visible
At the 22 July Talk Concert, co-heads Woohyung Lim and Honglak Lee presented an offer extending beyond the base model. Jungkyu Choi introduced the work agent, Data Foundry and an on-premise package supporting both GPU and NPU infrastructure. LG's subsequent event account reported more than 50,000 group office workers using ChatEXAONE and more than 5.1 million downloads across the model series. Lim connected those developments to applying the research in industry.
FriendliAI chief executive Byeonggon Jeon demonstrated EXAONE on stage. FuriosaAI chief executive Joonho Baek discussed the inference work behind the on-premise offering, while LSEG's Arman Sahovic described the jointly developing Master Score financial-information project. Honglak Lee also introduced a visual-language model for complex documents and images. The presentations brought together the people building hosted access, local infrastructure and specialist applications.
Choi's Data Foundry addressed another part of the enterprise job: producing specialist training data while retaining control of sensitive information. LG described trials with group affiliates and public institutions. Its on-premise package let companies choose features and configure an installation within their own networks. The talk moved from the language model itself to the data, hardware and operating environment needed to use it.
The Model Continued Into Document and Image Work
LG kept a research route alongside the managed service. The live 32B weights licence permits research and education, with a separate agreement required for commercial exploitation. The API supplied hosted execution, while the July event added configurable local deployment. These were the practical formats through which LG was making its model family available.
The public ChatEXAONE beta introduced around the event has since concluded, according to LG's current event page. The research programme continued: on 9 April 2026, LG's EXAONE 4.5 report described training a 1.2B vision encoder and joining it to the 4.0 32B language backbone, with context extended to 256K tokens. The work carried the earlier model into reading images and technical documents. July's combination of reasoning, hosted access and local infrastructure had become part of a continuing enterprise product line.
DSML research · 8 October 2026 · Revised 31 October 2026



