Transformers adds native GGUF support, Next.js 16.3 cuts dev memory 90%

Transformers now loads quantized GGUF checkpoints directly via from_pretrained() with no llama.cpp bridge, matching llama-bench performance on Apple Silicon. Next.js 16.3 reduces dev memory from ~20 GB to ~2 GB via Turbopack disk caching. Unsloth Dynamic 3.0 improves quantization accuracy by 10%, but 1-bit quants still break agentic workloads.

• Transformers loads GGUF natively; Q4_K_M shrinks 8.42 GB model to 2.74 GB
• Next.js 16.3 cuts dev memory from ~20 GB to ~2 GB
• Unsloth Dynamic 3.0 boosts quantization accuracy 10% without retraining
• Miri leaks secrets into cached target/, exposed to fork PRs