Options for using multiple models

[BradKML]

Other than agentics and mathematics, usually FOSS is better. Ideally OpenEvolve for proof-of-concept work should heavily bias Qwen3-32B or Qwen3-30B-A3B, and only use larger Qwen coder for heavyweight development SomeOddCodeGuy/WilmerAI#26 (reply in thread)

This line of thought makes me think of a few things:

1. How do we find better reinforcement learning algorithms? What comes after "Dr. GRPO" and RLVR? Implicit rewards? Maybe self-play? Open-ended ("free range") research optimization #197 What about "Intuitor"? https://github.com/sunblaze-ucb/Intuitor
2. How do we transfer the skills of one extremely unhostable model into open weight industrial models and then to something more consumer-friendly? (yes I am essentially asking for a multi-tiered distillation solution)
3. Since RouteLLM only interpolates between 2 models, and WilmerAI only routes through task type, what are the possible ways of handling both at the same time, and using more than 2 models per task type? https://github.com/lm-sys/RouteLLM