Three stories from the same week, read together: the one-year review of Switzerland's open model Apertus, DeepSeek's first multimodal release under clean MIT, and an open model specializing in medicine. What actually matters for anyone weighing an architecture today.
In nine days, five open-weight releases confirmed that the capability gap with closed models is closing faster than expected — and also showed, with a concrete case, why yesterday's license is no guarantee for tomorrow's.
In a single week: MIT weights already downloadable, a freshly published revenue-share license, a model that finds critical vulnerabilities and keeps its weights locked, and a fourth model that doesn't even have a name. Here's what changes for anyone deciding today.
Meta reverses its closed strategy, Nvidia ships its very first open source model. Same week, same direction: small models, licenses that hold up to scrutiny, built to run on a single GPU.
In the same week, Alibaba announces an "open" model with no weights and no license, DeepSeek publishes its own under MIT, and the European Union picks who will build its sovereign open model. Three very different degrees of openness.
Two stories from the same week, one from Brussels and one from Washington, say more about where open-model regulation is heading than any benchmark leaderboard. Here's what changes for companies already running them in Switzerland.
Kimi K3's full weights — 2.8 trillion parameters, the largest open-weight release ever — went live today. But the choice that matters for a company isn't which model to download: it's where to run it.
Two weeks after our last round-up, the open-weight leaderboard has already flipped twice. Here's what Kimi K3 and Inkling are, their licenses, and why chasing the podium isn't the useful part.
In the first six months of 2026 the open-weight leaderboard changed more times than most people could keep up with. Here's what actually matters for a company, beyond the benchmark noise.
The biggest model isn't always the best choice. Optimization techniques often deliver the same perceived quality at a fraction of the hardware cost.
RAG and fine-tuning are often presented as equivalent alternatives. They aren't: they solve different problems, and choosing the wrong one costs months of work.
Adopting an AI tool without checking where the data ends up is the fastest way to turn a productivity gain into a compliance problem. Here's what to check first.
A closed model and an open-weight model aren't two variants of the same product: they change who sees your data and who controls the system. Here's how to decide, case by case.
AI agents promise to automate entire workflows. On open models, today, some use cases work well in production — others don't, and it's better to know which before you build on top of them.
The word "open" gets used very loosely in the AI world. The licenses behind the models, however, differ significantly — and reading them before adoption is not optional.
"Do we need a supercomputer?" is the question that stalls more AI projects than any other before they even start. The answer is almost always more modest than feared.
"Open model" isn't a single category: behind the term are very different families. Here's how to find your way without letting benchmarks alone guide you.