Qwen 3.8 27B tops Hugging Face — and runs on one local GPU
Apache-2.0, 27B, native vision — a 4-bit build fits ~17 GB, GGUF and Mac forks are live, and one tester had it reverse-engineer a license check in 30 min.
Copy markdown
#1 trending, and it's Apache-2.0
Qwen 3.8 27B is the top trending model on Hugging Face right now — a 27B native vision-language model under a permissive Apache-2.0 license, with 262K context (extensible to 1M via RoPE) and ~2.3M downloads in the last month. You can fine-tune it, self-host it, and ship it commercially with no API key.
Fits in ~17 GB — one card or a Mac
Quantized to 4-bit it drops to roughly 17 GB, and unsloth's GGUF build is already live — so it runs on a single 24 GB GPU or a unified-memory Mac. One hands-on clocked about 50 tokens/sec on a 128 GB GB10 workstation. That means fully local, private, offline inference instead of metered API calls.
It reverse-engineered a license check in 30 min
In a hands-on write-up, XDA reports the model deconstructed a commercial app's license-verification system and produced a working bypass in about half an hour — running entirely on local hardware. Agentic reverse-engineering you'd once have handed a frontier API now runs on the box next to you, with the dual-use caveats that implies.
Community forks are already shipping
Within days, uncensored MLX builds for Apple Silicon, FP8 variants for faster inference, and 'obliterated' derivatives are trending right alongside the base weights. Pick the build that matches your hardware and your guardrail tolerance — the ecosystem moved before the model was a week old.
Why now: the frontier tax
Builders are increasingly routing work to cheaper models as premium tiers strain budgets — Simon Willison, citing the Ramp AI Index, notes teams offloading from the priciest frontier models. A capable local 27B under Apache-2.0 turns that cost dodge from a workaround into a real default.