Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – August 20, 2026
The Local Inference Speed Race
In this episode:
The Local Inference Speed Race — A developer pushed Qwen3.8-27B to 138 tokens per second on a single RTX 3090 using DFlash2 speculative decoding and lookup-augmented drafting, while llama.cpp added adaptive multi-token prediction and NInfer claimed 350K-token context on one 4090.
Quantization and Intelligence per Watt — Unsloth released Dynamic v3.0 GGUFs claiming over 10% accuracy gains at the same file size using KL-divergence evaluation, while Liquid AI shipped QAD distillation-based quantization recovering 97% of lost accuracy at…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.