2026 vLLM Korea Meetup

The 2026 vLLM Korea Meetup, hosted by the vLLM KR Community and supported by Rebellions, SqueezeBits, Red Hat APAC, and PyTorch Korea, was held in Seoul on April 2. On-site participation was high, and the post-event survey response rate reached about 75%, showing that both technical sharing and community experience worked strongly.

The event's core message was clear. LLM serving is no longer a matter of choosing a particular framework, but is turning into an infrastructure problem of stably operating diverse models and hardware.

In the opening session, Hongseok Kim of Rebellions and Li Ming of Red Hat APAC introduced the latest trends in the vLLM project and community.

• Over the past 6 months, the vLLM KR community has built structure through Steering Group-centered operations, regular meetups, and hands-on workshops.
• Technically, the full architectural migration from v0 to v1 is underway, strengthening codebase simplification and modularity.
• Internally, async scheduling and Model Runner improvements were made, and features such as streaming API, semantic router, and vLLM-Omni have rapidly increased.

In the hardware integration part, the direction for combining vLLM with AI accelerators was presented. Rebellions is developing the vllm-rbln plugin to bring its own NPU into the vLLM ecosystem, and already supports paged attention and continuous batching. Going forward, they plan to add speculative decoding, distributed KV cache, and prefill/decode disaggregation, with an eye toward next-generation NPU-based large-scale inference clusters like Rebel100™.

In the same session, Li Ming introduced vllm-playground to lower the entry barrier. It eases the initial user experience—which had been difficult due to over 140 configuration parameters—with a GUI, enables experimentation on CPU and macOS environments, and also provides performance visualization.

The afternoon session covered the present and future surrounding the vLLM production stack. The message was singular: vLLM is continuously absorbing the functionality and scalability needed for real production environments, going beyond a simple model serving engine.

From the middle of the event onward, practical case studies were presented across two tracks.

• Track 1: vLLM with Open Source

  • Juho Lee of XCENA defined LLM serving essentially as a cluster efficiency problem, explaining that how KV cache is stored and reused determines both performance and cost simultaneously.
  • He presented KV cache tiering and routing via LMCache, and an approach using CXL memory as a large-capacity cache extension layer.
  • Inseo Song of Upstage shared how Chat Templates, token-level state parsing, parsers, and logits processors are used in the process of turning open-source models into actual services. The key point was that "serving stably" is a harder problem than "building a good model."

• Track 2: vLLM in Business

  • Sungsu Kim of Samsung Electronics emphasized that security is the top priority in enterprise environments handling sensitive data. In situations where external SaaS models cannot be used, they built a private LLM API based on internal GPU infrastructure and an air-gapped closed network, connecting it to OpenWebUI, OpenAI-compatible APIs, Dify, and Claude Code, with over 4,000 users currently using it.
  • Jaeeun Gil of NAVER Cloud shared the serving case of HyperCLOVA Omni. Since omni-modal models handling text, images, and audio together are structurally heterogeneous, efficient serving is difficult with existing methods, so they applied a disaggregated serving architecture. Analysis revealed that the vision decoder was the bottleneck accounting for most of the overall latency, and they achieved over 3x performance improvement through sequence parallelism and kernel optimization.

Finally, attendees highly praised the event's technical focus and the density and vitality of the vLLM ecosystem. The trend of hardware vendors, cloud providers, AI service companies, and end users all building their strategies around vLLM has become clear, and this meetup vividly demonstrated that shift.