Deploying Large Language Models (LLMs) is becoming a core skill for developers, but most people still rely on “black-box” APIs. In reality, choosing the right serving engine can drastically change latency, throughput, cost, and user experience.

In this session, I’ll take the audience on a practical, developer-oriented journey comparing Ollama and vLLM, the two most popular open-source solutions for running LLMs locally and in production.

We will break down the concepts simply, in Darija, and based on real benchmarks:

+ What we will learn
-How LLM serving actually works (tokenization, batching, memory planning, GPU scheduling)
-Why vLLM is extremely fast: Continuous batching, PagedAttention, optimized sampling
-Why Ollama is extremely easy: UX-first design, model packaging, Docker-like simplicity
-When to choose Ollama (local dev, prototyping, on-device apps, offline assistants)
-When to choose vLLM (APIs, high-throughput apps, RAG, production environments)
-Live architecture comparison: performance, GPU/CPU usage, limits, ecosystem
-Practical deployment demo: serving the same model using both tools
-Real-world lessons from building AI educational tools and agents

+ Why this talk matters

-Most developers in Morocco use LLMs but don’t know:

how model serving really works,
how to reduce latency,
or which tool fits their use case.

This session gives the community actionable, Darija-friendly knowledge, enabling them to build faster, cheaper, more scalable AI apps.

+ Target audience

Beginner → Advanced developers interested in:
-AI & LLMs
-backend & infra
-building AI products
-RAG, agents, inference optimization

By the end, the audience will leave with a clear mental map of how to run LLMs locally and in production, and the confidence to pick the right stack for their next AI project.