ETRI open-sources TANGO 2, a no-code tool for fine-tuning and deploying LLMs on industry data

ETRI has worked on TANGO since 2021. The first version handled computer vision tasks such as image recognition, and TANGO 2 extends it to generative AI and LLMs[1]. Its core is an LLMOps platform that runs on a Kubernetes cluster and exposes the whole LLM and VLM workflow through a web UI[2].

The README lists five working areas[2]:

• Data: upload, preprocessing, filtering, generation, human and VLM review, and versioned dataset releases.
• Model: LoRA fine-tuning from Hugging Face or platform models, live training graphs, checkpoints and GGUF quantisation.
• Deployment: cluster serving on vLLM, llama.cpp, TGI, Ollama, Triton or TensorRT-LLM, plus on-device, standalone package and external API options.
• Inference: chat and batch inference against a running deployment, image input for VLMs and prompt optimisation.
• Evaluation: lm-evaluation-harness benchmarks, LLM judging with the Korean LogicKor benchmark, intent and slot scoring, and expert review reports.

A separate runtime layer handles where the model ends up, from cloud and on-premises servers down to Jetson Orin, Raspberry Pi and PCs, with real-time scheduling, failover and monitoring for edge devices[2]. The latest tagged release, tango2-26.11, landed on GitHub on 5 November with VLM demos, an OpenTelemetry tracing stack and a restructured LLMOps codebase[3].

The repository ships with three industry trials alongside the tool itself[2]. In shipbuilding, Avnautics is testing an AI navigation aid that reads data from a vessel's sensors. In farming, Sure Data Lab is building a generative AI service trained on expert growers' knowledge. In medicine, Seoul National University Hospital has developed automatic radiology report generation from chest CT scans, with a cardiopulmonary and vascular disease prediction trial planned with Seoul National University Bundang Hospital[1].

On the commercial side, ETRI says the LLMOps tool is being linked to Acryl's Jonathan product, and agent AI company Persona AI licensed TANGO technology in April for its GEN STATION platform[1].

"Open source" here comes with conditions. ETRI's licence agreement applies GPLv3 or later to open-source use, but any commercial use of the code or derived work, including free distribution supported by advertising, falls under a separate ETRI commercial licence[4]. Teams planning a product on top of it should talk to ETRI first.

Most LLMOps platforms assume an ML engineer at the keyboard. TANGO 2 is a government-funded attempt to push fine-tuning and edge deployment down to domain experts, and it comes with working industry examples rather than just a framework. ETRI will walk through the tool at the 5th TANGO Open Seminar in Seoul on 11 November and is seeking an ITU-T standard for no-code LLMOps, with final approval targeted for May 2027[1].