AGI Dreams Podcast – October 26, 2026

AGI Dreams – Open, Uncensored, & Local - AI News Digest · Oct 26, 2026 · 22m · transcript

Paged Experts and Tiny Quants

In this episode:

Paged Experts and Tiny Quants — GreenBitAI's 4-bit paged DeepSeek V4.1-Flash enables 406 GiB models on memory-limited machines, while llama.cpp builds push extreme quantization to run large coding models on 8 GB GPUs. Runtime abliteration and ultra-compact quants keep driving local inference toward lower hardware floors.

Unlocked Silicon — A Linux kernel module unlocks hidden FP16 tensor core performance on NVIDIA Turing GPUs, measuring up to 160% of official peak TFLOPS. Community reports extend the technique across three GPU…