<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Posts on Noureddine RAMDI</title><link>https://ramdi.fr/post/</link><description>Recent content in Posts on Noureddine RAMDI</description><generator>Hugo</generator><language>en</language><lastBuildDate>Sat, 23 May 2026 20:41:27 +0000</lastBuildDate><atom:link href="https://ramdi.fr/post/index.xml" rel="self" type="application/rss+xml"/><item><title>Hunting Tokens/sec: 4 LLM Backends, 1 Hard Ceiling (Part 2/4)</title><link>https://ramdi.fr/post/ai-llm/local-llm-tokens-per-second-benchmark/</link><pubDate>Tue, 28 Apr 2026 00:00:00 +0000</pubDate><guid>https://ramdi.fr/post/ai-llm/local-llm-tokens-per-second-benchmark/</guid><description>Part 2 of 4: a benchmark journal across nixpkgs llama.cpp, upstream master, and ik_llama.cpp on Qwen3.6-27B. Six hours, four backends, all converging at 66 tok/s — and the physical reason why.</description></item><item><title>Speculative Decoding Meets Hybrid SSM: Why It Breaks (Part 3/4)</title><link>https://ramdi.fr/post/ai-llm/local-llm-speculative-decoding-hybrid-ssm/</link><pubDate>Tue, 28 Apr 2026 00:00:00 +0000</pubDate><guid>https://ramdi.fr/post/ai-llm/local-llm-speculative-decoding-hybrid-ssm/</guid><description>Part 3 of 4: a deep-dive into why speculative decoding silently breaks (or runs anti-economically) on hybrid attention+SSM architectures like Qwen3.6, Mamba-2, and RWKV — and what would need to change upstream to fix it.</description></item><item><title>The NixOS Setup for llama.cpp: Declarative and Reproducible (Part 4/4)</title><link>https://ramdi.fr/post/ai-llm/local-llm-nixos-llama-server-module/</link><pubDate>Tue, 28 Apr 2026 00:00:00 +0000</pubDate><guid>https://ramdi.fr/post/ai-llm/local-llm-nixos-llama-server-module/</guid><description>Part 4 of 4: the actual NixOS module, llama-pull helper, claude-code-router wiring, and one-line workflow for switching models. Five Nix files for a complete, isolated, rollback-able local LLM service.</description></item><item><title>Why I Serve Qwen3.6 Locally on My RTX 5090 (Part 1/4)</title><link>https://ramdi.fr/post/ai-llm/local-llm-rtx5090-why-nixos/</link><pubDate>Tue, 28 Apr 2026 00:00:00 +0000</pubDate><guid>https://ramdi.fr/post/ai-llm/local-llm-rtx5090-why-nixos/</guid><description>Part 1 of 4: motivation, hardware, and stack choices for serving Qwen3.6-27B locally on a 32 GB consumer GPU with NixOS, before any benchmarks or trade-offs kick in.</description></item><item><title>Getting Started with Nix Flakes</title><link>https://ramdi.fr/post/nixos/getting-started-with-nix-flakes/</link><pubDate>Sun, 12 Apr 2026 00:00:00 +0000</pubDate><guid>https://ramdi.fr/post/nixos/getting-started-with-nix-flakes/</guid><description>A practical introduction to Nix Flakes — the modern way to manage NixOS configurations and development environments.</description></item><item><title>Building AI Agents with Claude Code</title><link>https://ramdi.fr/post/ai-llm/building-ai-agents-with-claude/</link><pubDate>Sat, 11 Apr 2026 00:00:00 +0000</pubDate><guid>https://ramdi.fr/post/ai-llm/building-ai-agents-with-claude/</guid><description>How to leverage Claude Code to build autonomous AI agents that publish content, review code, and manage workflows.</description></item></channel></rss>