Local models and open harnesses are the hedge against AI vendors

Liquid AI's LFM2.5-24B-A2B claims 184 tokens/s on an Apple M5 Max and 21K output tokens/s on a single B300. One commenter reported 200 to 320 tok/s on a 5070 Ti with 131k context and no tuning. OpenCode 2.0 shipped to warm reviews, and an Ask HN on replacing Claude or GPT with a local model drew real field reports. Someone runs DeepSeek V4 Flash on 2x RTX Pro 6000 at 160 tok/s, and another says Qwen 3.6 27B dense feels roughly like Haiku 4.5.

The sharpest context is the thread where Anthropic stops letting Claude Code subscriptions run through OpenClaw. Anthropic called third-party tools an outsized strain and offered a one-time credit. Commenters pointed out that this is the capacity they paid for. One says they hit rate limits within 1:45 on afternoons. Here is the pattern: every time a vendor tightens a subscription, the case for open harnesses and local weights gets stronger.

The counterpoint is honest. Several people tried Gemma 4 or similar on an M4 and found the speed and quality well below cloud offerings. Speed is no longer the main gap. Intelligence per task is, and nobody in the thread claims local matches the frontier for hard work.