Alibaba released Qwen3.8-27B under Apache 2.0 — a dense 27B model with native vision and a 262K context window, running locally through Ollama or vLLM.
In favour
Fully open, commercial use allowed — no per-token fees, nothing leaves your box.
Fits a 24 GB GPU (≈13 GB quantised); the vision head reads logs, UIs and diagrams.
Worth watching
Its own benchmark sheet compares it to last-generation frontier models, not today's top tier — impressive, not a frontier-killer.
On integrated / shared-memory GPUs it is bandwidth-bound and slow; a discrete GPU makes it comfortable.
Our takefor sovereignty the question isn't frontier parity, it's "last year's frontier that you own and run for €0 a token." For a growing number of teams, that trade is now worth making.