64GB no longer a dead zone for local AI agents

A user reports running Qwen3 27B for agentic coding on a 64GB M4 Max, using about 48GB at 256k context. A 93% prompt cache hit rate means each turn processes only ~1k new tokens instead of 160k, avoiding a 20-minute prefill. With agents, context costs as much as weights, making 64GB a sweet spot.

• User tests LFM2.5-24B-A2B, finds local performance poor2026-10-22
• Transformers now runs llama.cpp quants2026-09-22
• Study: Pretraining Poisoning Can Plant Agentic Backdoors2026-11-02
• Developer Measures Output Damage From Engine Speed Flags2026-11-01
• Developer runs Qwen 3.8 Flash Next on an RTX 3060 at 22 tok/s2026-10-31