Unsloth Dynamic 3.0 GGUFs run Qwen3.8-27B locally in 6.2GB
New dynamic quants beat rival GGUFs by ~10% at the same size — plus Vercel open-sources fx, a Zig coding agent, and Claude Code ships a Concise mode.
Copy markdown
27B on a laptop, at 6.2GB
The new UD-IQ1_S quant shrinks Qwen3.8-27B to ~6.2GB — 89% smaller — while holding ~72% top-1 accuracy; UD-Q2_K_XL lands at 9.83GB. A 27B model now fits on a 16GB machine.
~10% more accurate at the same size
On KL-divergence and a new Divergence-300 trajectory test, UD-3.0 files hit up to 10% higher top-1 accuracy than other providers' GGUFs at identical disk sizes — all pure post-training, no QAT.
These are new files — re-download
Dynamic 3.0 isn't a toggle in LM Studio or llama.cpp; it's a fresh set of GGUFs. If you pulled Qwen3.8 during its 5M+ downloads this week, grab the v3 files to get the gains.
Tuned for coding, chat, multilingual
The gains come from a higher-quality imatrix calibration set aimed at agentic coding, chat, and multilingual use, plus smarter per-layer quant selection. Runs in llama.cpp and Unsloth Desktop.
Elsewhere: Vercel open-sources fx, a Zig coding agent
Vercel Labs released fx, a tiny Unix-style coding-agent harness and CLI written in Zig — fast and built to embed in bigger systems. Clone it and roll your own agent this weekend.
Elsewhere: Claude Code 2.1.237 adds a Concise mode
A new built-in "Concise" output style leads with results and skips preamble; the same release fixes prompt caching for anyone on an LLM gateway or custom base URL.