Mminhleducinminhleduc.hashnode.dev·3d ago · 15 min readKimi K3 has 2.8T parameters. Moving them was the hard part.Every article about Kimi K3 leads with the same number. 2.8 trillion parameters, the largest open-weight model anyone has shipped. It's a real number and it's the least interesting thing in the techni00
Mminhleducinminhleduc.hashnode.dev·3d ago · 12 min readHow Kimi K3 reads a million tokensStandard attention has one genuinely magical property: any token can look directly at any earlier token, at full fidelity, no matter how far back it sits. Nothing is summarised. Nothing is forgotten. 00
Mminhleducinminhleduc.hashnode.dev·5d ago · 12 min read5 Tiny Language Models for Tool Calling (Part 3)The benchmark nobody quotes when they tell you small models are the future. NVIDIA Research published a position paper called "Small Language Models are the Future of Agentic AI". I believe the argume00
Mminhleducinminhleduc.hashnode.dev·5d ago · 10 min read5 Tiny Language Models You Need to Know in 2026 (Part 2)I wrote Part 1 in October 2024. The subtitle was "Tiny is a new large!" and at the time that felt like a slightly cheeky thing to say. It doesn't anymore. Google now ships a model that runs in 1.1 GB 00
Mminhleducinminhleduc.hashnode.dev·Aug 1 · 9 min readFrom Retrieval to Memory: The Graph Your Agent Writes ToPart 3 of 3 — Knowledge Graphs for AI Engineers. Part 1 covered why vector search hits a ceiling. Part 2 covered building the graph and what it costs. This one is about the shift that actually changed10
