Unsloth Dynamic 3.0 GGUFs: +10% top-1% accuracy, same size
The new calibration is tuned for agentic coding and runs on llama.cpp today — plus Claude Code's Concise mode and OneCLI's agent firewall.
Copy markdown
Same file size, more of the model
Dynamic 3.0 rebuilds each quant with a better calibration set and smarter layer selection, landing up to 10% closer to full BF16 top-1% accuracy than competing GGUFs at the same disk size — pure post-training quantization, no retraining.
Run Qwen3.8-27B on a light rig
The first drop is Qwen3.8-27B: the UD-IQ1_S build is 6.2GB — an 89% cut — and still keeps ~72% top-1% accuracy, while UD-Q2_K_XL at 9.83GB beats the next-best quant by 8%. It loads straight into llama.cpp.
Tuned for coding agents, not perplexity
The calibration data was refined for agentic coding, chat, and multilingual use, so the quality you keep maps to real coding-loop work rather than a generic perplexity score. Its Divergence-300 metric scores 32-token trajectories, not single-token guesses.
Elsewhere: Claude Code 2.1.237 adds Concise mode
Today's Claude Code ships a built-in 'Concise' output style that leads with results and skips preamble — set it under Output style in /config — and fixes prompt caching for sessions behind an LLM gateway or custom base URL.
Elsewhere: OneCLI wraps any agent in a firewall
OneCLI (YC S26) launched an open-source sandboxed harness that boxes Claude Code, Codex, or any agent with per-employee isolation and network controls — clone it to give a team secured personal agents this weekend.