Unsloth v0.1.803-beta adds experimental auto compaction that keeps evicted turns searchable past the context limit

GitHub·medium signalXBlueskyLinkedInCopy link

Released 2026-08-25 with 170+ PRs, the update lets long local chats continue past a model's context limit by rolling older turns out of active context into fresh context epochs rather than permanently trimming replies, with evicted conversations still searchable and improved fact recall. It also fixes MLX and Mac runtimes that did not run correctly, adds keyless LAN access without Cloudflare links, and supports custom llama.cpp builds with toggles for cache RAM, mmap, mlock and speculative-decoding KV cache. Unsloth Dynamic v3.0 GGUFs for Qwen3.8-27B are claimed to deliver >10% higher top-1 accuracy than competing quantizations.