Near-frontier models became downloadable and self-hostable. Access to strong AI stopped depending on a single vendor.
About this eventAn open-weight sparse mixture of experts in which only part of the parameters runs for each token.
Mixtral contains several expert blocks, while a router activates only two per layer and token. This increases total capacity without proportional inference cost; open weights and the Apache 2.0 licence allowed researchers and companies to deploy and adapt the system themselves.
Source 1: Mistral AI · техническое описаниеSource 2: Mistral AI · карточка моделиAbout this eventThe first open model comparable in scale to the closed frontier.
About this eventStrong reasoning at a radically lower training cost.
About this eventA dense open family for every hardware budget.
About this eventMoonshot AI published Kimi K3 weights with launch instructions for Transformers, vLLM and SGLang. This is an infrastructure-scale release, not a model for an ordinary home PC.
Published files and instructions for several serving runtimes move Kimi K3 from announcement to an available open-weights release. At its stated 2.8T-parameter scale, full inference requires serious cluster infrastructure. The licence must be read in the model card: open weights does not itself mean Apache or MIT.
Source 1: Moonshot AI · Kimi K3 model cardSource 2: vLLM · day-0 supportAbout this eventOpen models trail closed ones by roughly six to twelve months, and the gap keeps opening and closing.
About this eventRegulators are debating which capabilities must not be published irreversibly. Nobody has clear lines yet.
Primary material behind this dossier: papers, lab publications and official reports.