Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – September 15, 2026
Big Models on Small Iron
In this episode:
Big Models on Small Iron — A one-bit 27B-parameter model now runs at 25-30 tokens per second in a browser on a 6 GB GPU via WebGPU, while other setups push DeepSeek V4.1 Flash on CPU-only and Qwen3-Flash-Next on dual 3090s, showing consumer hardware closing the gap on inference.
Context, Memory, and Migration — Spomin performs live KV-cache compaction by replacing transcript chunks with summaries mid-session, fitting 500k+ tokens of material into 180k resident. Separately, embedflow proposes migrating embedding models via reranking rather…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.