backend.ai

Lablup is now a Participating Sponsor of the Python Software Foundation (PSF).

In January 2026, the Upstage consortium that Lablup is part of successfully passed the Phase 1 evaluation for the Korean government's Sovereign AI Foundation Model project. This initiative aims to protect national AI sovereignty by having the government provide support for GPUs, data, and talent development, while the private sector actively leverages these resources to develop frontier-grade AI foundation models. We sat down with team members from Upstage and Lablup to hear the behind-the-scenes story of our Phase 1 journey.

This article covers the major changes in Backend.AI FastTrack 3 25.18.

Backend.AI 25.15 LTS is now officially available. This release brings comprehensive system-level optimization and user experience improvements, reinforcing the platform’s reliability and scalability for large-scale AI model training, deployment, and research.

We're excited to announce Backend.AI 25.6, the first Long Term Support (LTS) release of 2025. This update brings significant improvements to system monitoring, audit logging, and model service auto-scaling, making operations more convenient than ever.

To write Storybook stories for 50+ BAI components in the Backend.AI WebUI, I started by setting up the infrastructure— i18n, theming, and branding — then upgraded to Storybook v10 and merged two instances into one. An automation pipeline combining a 1,000-line guideline, Claude-based story generation, and GitHub Actions CI checks kept quality consistent from PR creation through deployment. The key takeaway: build the foundation and automation first, and when working with AI, the human role shifts from writing code to defining and refining standards.

Lablup is the infrastructure partner in the Upstage consortium for Korea's national Sovereign AI Foundation Model project. We're pre-training Upstage's Solar Open (a 102B MoE model) on SKT's "HAEIN" cluster, using around 500 of its 1,000+ NVIDIA B200 GPUs. When you operate a cluster at this scale, one reality hits you fast: failures are not a matter of if but when. To reduce downtime from these inevitable failures, we've been building a proactive failure detection system that aims to predict which GPUs are likely to fail and preemptively swap them out. This post covers the first step: what data we collect from the cluster, how we collect it, and what early warning signs we've actually observed in real failures.

DGX Spark is a desktop AI supercomputer that packs 128GB of unified memory and 1 PFLOP-class Grace Blackwell (GB10) performance into a palm-sized box. However, its internal GPU belongs to the SM12x series, distinct from the data center-grade Blackwell (SM100). This creates a subtle architectural gap: the latest LLM stacks, heavily reliant on MLA·DSA-specific kernels like GLM-5, "Blackwell support" alone doesn't guarantee immediate compatibility. This creates a subtle architectural gap requiring separate code management for Hopper, data center Blackwell, and consumer Blackwell. The engineering team examines Spark, which is based on Blackwell but features a slightly different architecture.