This session is a technical walkthrough of what it takes to run Open Frontier AI in production: inference, SFT, custom speculative decoding, and dedicated endpoints, on Nebius Token Factory.

​What to expect

​A technical presentation and demo of how to run full-stack AI Engineering Pipeline with Nebius Token Factory.

1. Choosing your model stack — open weights vs. proprietary APIs

​A decision framework, not a vendor pitch. When open-weight models win on cost, latency, data control, or flexibility, and when they don't. What Kimi K3 changes about that calculus now that a 2.8T MoE model ships at open weights and scores 57 on Artificial Analysis' Intelligence Index, just two points behind GPT-5.6 Sol.

​2. Building on Kimi K3

​The architecture that makes it work: 2.8T total parameters activating only 16 of 896 experts per token, Kimi Delta Attention, native vision, a full 1M-token window. Where K3 outperforms alternatives on long-horizon coding and agentic tasks, where it doesn't, and integration patterns through Token Factory's OpenAI-compatible API.

​3. Full-stack AI engineering — demoed across the whole pipeline

• ​Inference — Fast vs. Base flavors, time-to-first-token under load, multi-region routing, and when to pay for which
• ​Post-training (SFT) — LoRA and full fine-tunes on your own data, and promoting a tuned model straight into production
• Custom speculative decoding — training your own draft model against your workload's actual traffic shape, and what that buys you on latency
• Dedicated endpoints — performance isolation, reserved capacity, autoscaling, and the point in your growth curve where the switch pays for itself

​4. Open floor

​Talk to us and your peers about the hardest AI Problems you are trying to solve.

​Who this is for

​CTOs, VPs of Engineering, AI Architects, AI FDEs and senior AI engineers at growth-stage startups building AI-native or AI-augmented products — typically Series A and beyond.