llama-cpp 2 Hitting the SmolLM2 ceiling May 1, 2026 Rolling out 4 new providers (2 local, 2 cloud) — and the AI SDK bug we hit Apr 28, 2026