PLI 4.0.0 pre-release targets M1 Ultra

PLI is impressive hardware-specific engineering, but its headline performance is closer to a carefully tuned appliance than a general-purpose runtime. The payoff is strongest for long-context local agents; the tradeoff is narrow validation and dependence on private Apple APIs.

• –Public benchmarks show PLI outperforming Kapitsa, llama.cpp, and mlx-lm on long prompt processing, while Kapitsa remains faster on short-prompt generation.
• –Hybrid DFlash2 and MTP drafting, checkpoint reuse, reply retention, and disk spilling are designed specifically for sustained agent sessions.
• –Existing OpenAI and Anthropic clients can connect without custom adapters, making PLI practical for local coding workflows.
• –The server is Apple Silicon-only, validated on one M1 Ultra and one model family, with roughly 36 GB of wired memory and potentially large disk snapshots.
• –Neural Engine integration could break after macOS updates, so this is promising experimental infrastructure rather than a drop-in production default.