A dense 27B ships under Apache 2.0 with native video
Native vision and video in a dense 27-billion-parameter model, Apache 2.0, 262K context extensible toward 1M, with thinking control. Video in an open dense model at this size is the new part.
A dense 27-billion-parameter multimodal model has been released under Apache 2.0 with native vision and video, a 262K context window extensible toward 1M, and flexible thinking control.
Video is the part that has not been available at this size under these terms. Image understanding in open models is now ordinary; video means temporal reasoning, and temporal reasoning at 27B on permissive terms puts a class of application — surveillance review, inspection footage, sports and clinical video — inside reach of a team with one server.
Dense rather than sparse matters here too. A dense 27B has predictable latency and memory, which is what you need when the workload runs on fixed hardware rather than in an elastic cloud. Sparse models win on cost per token at scale and lose on being possible to deploy on a box you own.
The thinking control lines up with a pattern visible across the month. Reasoning effort is becoming a caller-set parameter rather than a property of the model, and it is now appearing in open releases as well as hosted APIs.
Digital Applied — AI Model Releases: August 2026 Tracker and Dated Ledger → · Punku — Best Open-Source AI Models 2026: The Data Comparison →