glasswing · live models

Two small language models running on an inference engine written from scratch in Rust, served from a single-node Kubernetes cluster at home. Slow on purpose: no GPU, no batching, just the CPU. How it works.

No login and no logging of what you write, so there are limits: a few requests a minute per visitor, up to 256 tokens per answer, two answers at a time across everyone. SmolLM2-135M is Hugging Face's (Apache-2.0) and is tiny: expect confident nonsense. glasswing only knows children's stories.