Interconnects Analysis Maps the Open-Model Power Balance After a Year of Shifting Releases

Nathan Lambert's Interconnects newsletter published an analysis this week titled "The current balance of power in open models," an attempt to map which labs and countries hold real leverage in open-weight AI after a year in which Meta, DeepSeek, Alibaba's Qwen team and Mistral repeatedly reshuffled the rankings. The core argument is that the open ecosystem should be read as a distribution of power rather than a single leaderboard, because who leads depends on the metric you pick. Raw benchmark scores, download counts, the number of derivative fine-tunes, and paid commercial deployments each point to different winners.

That distinction has stopped being academic. Over the past twelve months, open releases have closed much of the capability gap with closed frontier systems faster than most forecasters projected, and the pattern of who ships weights no longer fits the old assumption that only heavily funded Western labs can compete at the top. Chinese labs have made open publication a central strategy, releasing weights that developers anywhere can fine-tune and run on their own hardware without paying per-token API fees. The result is a market where a free model a few months behind the frontier is often good enough for production work.

Licensing has become as decisive as capability. If the strongest available weights carry restrictions on commercial use, or come from a jurisdiction regulators are scrutinizing, that changes what a startup can actually ship. Several recent releases have tightened or loosened their terms, and those changes have visibly redirected which models downstream teams adopt. Lambert's framing captures this by treating legal terms and geography as first-class variables in the power map, not footnotes to a benchmark table.

The practical stakes run in several directions at once. Companies deciding whether to keep paying for closed-model access now weigh that cost against a downloadable alternative that runs on private hardware with no per-query charge. Governments drafting AI rules still have to settle whether publishing open weights counts as a release, a publication, or a controlled export — a question with no settled answer in most jurisdictions. Safety and misuse researchers, meanwhile, need an accurate inventory of which models are genuinely in circulation rather than which ones made the most noise at launch.

What to watch is whether this distribution survives the next release cycle. If one lab pulls clearly ahead on capability while keeping its weights open, every downstream calculation shifts again and this analysis needs a rewrite. If the lead keeps trading among several players, the more durable finding is that no single lab controls the open ecosystem — which is the fact that matters most to anyone planning to build on it.