Deep Dive: LLM Quantization, part 3 - FP8, FP4

Two years after parts 1 (https://youtu.be/kw7S-3s50uk) and 2 (https://youtu.be/fXBBwCIA0Ds), the quantization landscape has changed completely. FP8 is the new default; the serving stack has absorbed quantization, and MoE models have broken the old assumptions. This video shows you what practitioners actually use today. In Part 3 of this series, I'm showing you exactly what to run, on which GPU, with which tool, and what breaks when you get it wrong. One model, start to finish: Arcee Trinity Mini (26B MoE, 3B active, 128 experts). What's covered: → Why FP8 is the new BF16 — essentially lossless, one flag, works everywhere → The three things people call "4-bit" (INT4 vs MXFP4 vs NVFP4) and why they're completely different → Serving with vLLM: BF16, FP8 on-the-fly, FP8 checkpoint, with verified H100 numbers → Why MoE models break standard quantization and how to fix it → Creating your own checkpoints with llm-compressor and NVIDIA ModelOpt → NVFP4 on Blackwell: real 4-bit tensor core math, what works and what doesn't yet → The FP4 deep dive: E2M1 bit layout, MXFP4 vs NVFP4 scale factors, worked error example → Decision tree: which quantization path for your hardware and use case *** Slides https://fr.slideshare.net/slideshow/advanced-quantization-techniques-for-large-language-models-in-2026-a4c8/286754686 *** Arcee Trinity Mini: 26B total parameters, 3B active per token, 128 experts, 128K context, Apache 2.0. https://huggingface.co/arcee-ai/Trinity-Mini NVFP4: https://huggingface.co/arcee-ai/Trinity-Mini-NVFP4 FP8: https://huggingface.co/arcee-ai/Trinity-Mini-FP8-Block #llm #quantization #fp8 #fp4 #nvfp4 #vllm #inference #nvidia #blackwell #moe #arcee