← Dashboard  /  News

2026-09-26

AI/Tech Digest 2026-09-26

AI & local LLM

1. Qwengram-0.8B: I transferred Qwen3.8 Flash-Next’s n-gram memory into Qwen3.5-0.8B — 5.05% lower validation perplexity [LAB]
   Demonstrates a specific method for improving small model performance via n-gram memory transfer.
   https://www.reddit.com/r/LocalLLaMA/comments/1wpvep4/qwengram08b_i_transferred_qwen38_flashnexts_ngram/ (reddit.com)
2. M5 Ultra 80Core GLM-5.3-Flash on DwarfStar Speeds [LAB]
   Provides concrete hardware performance benchmarks for running large models on high-memory Mac Studio setups.
   https://www.reddit.com/r/LocalLLaMA/comments/1wpkz0o/m5_ultra_80core_glm53flash_on_dwarfstar_speeds/ (reddit.com)