AGI Dreams Podcast – July 08, 2026

AGI Dreams – Open, Uncensored, & Local - AI News Digest · Jul 8, 2026 · 14m · transcript

Local AI Speed Wars

In this episode:

Local AI Speed Wars — DFlash speculative decoding lands in llama.cpp with a 4.44x speedup on long contexts, while Ollama 0.31 adds multi-token prediction for Gemma 4. AMD's Ryzen AI Halo dev kit offers 128GB unified memory for 70B models, and Huawei's Ascend 910C posts surprisingly competitive benchmarks.

Models Go Multimodal — NVIDIA's Nemotron Audex treats audio as first-class transformer tokens via a 30B MoE architecture with only 3B active parameters. Google's Gemma 4 delivers competitive multimodal performance under Apache 2.0, and…