Anubis benchmark analysis grows, 371 runs

This is the kind of update that makes a niche benchmark useful: more submissions, clearer metrics, and fewer apples-to-oranges comparisons.

• –371 runs and 218 models give the leaderboard more statistical weight, especially for Apple Silicon buyers comparing real-world performance.
• –Separating reasoning time from output throughput is a meaningful fix; mixed accounting can make a model look faster or slower than it really is.
• –The prefill/output split is the right direction for local inference benchmarking because prompt ingestion and generation bottlenecks are very different.
• –The dataset is becoming a tuning signal, not just a scoreboard, which makes it more valuable for model tweakers and Mac-local AI users.