Unsloth Dynamic 3.0 GGUFs: +10% accuracy at the same size
Better calibration and layer selection — no retraining — put a 27B model on a 16GB laptop; separately, Vercel open-sources a 6MB coding agent in Zig.
Copy markdown
10% more accuracy, same file size
Dynamic 3.0 rebuilds the imatrix calibration from a more diverse dataset and gets smarter about which layers stay high-precision, claiming up to +10% top-1% accuracy at matched disk size versus other GGUF providers. It's pure post-training quant — no QAT — scored on their Divergence-300 @32 eval.
Qwen3.8-27B fits in 6.2GB
The 27B model's 1-bit UD-IQ1_S build is ~6.2GB (roughly 89% smaller) and still holds ~72-77% top-1% accuracy; 4-bit lands at 16-19GB RAM, the practical sweet spot for a 24GB Mac or a 4090. These are new files, not a runtime toggle — re-download from Hugging Face to get them.
Elsewhere: Vercel open-sources 'fx', a 6MB agent
fx is a coding-agent harness and CLI written in Zig: a 6.4MB binary that cold-starts in ~10 microseconds, model-agnostic across local models, gateways, and provider APIs, and extended through skills, plugins, and MCPs. Apache-2.0, one-line curl install, small enough to embed in WASM or constrained runtimes.