Ollaya

No. Ollaya pulls open decision models and serves TypeSafe's /v1/systemone and /v1/decisions, so the official SDK works by pointing TYPESAFE_BASE_URL at http://localhost:11435. Ollama 0.35 also serves two decision models, Nimble and Tev1, through the same wire format, and the project FAQ compares the two directly.

Close, not equal. The recommended winnow:e4b scores 0.722 on typed decisions against Jev's published 0.738, at 89 ms for five questions on an RTX 4090. Without an NVIDIA GPU the README points you at laya, a 421M encoder that answers in a fraction of a second on CPU; the larger winnow sibling scores 0.702.

Weights stay with their authors. Ollaya publishes only small ONNX graphs of about 3 MB each, which read the original files from the author's Hugging Face repository pinned to a commit and verified by sha256; models whose authors ship GGUF, such as winnow, jevk5 and jeb, run that file on llama.cpp. It never re-hosts weights, and each model keeps its own license.

No, they are self-measured on the project's own GPUs and CPUs, with raw data published. What makes them more than marketing is the parity check: before a model ships, its runtime is compared question by question against the authors' code or llama.cpp on each device, and fp32 exports matched the PyTorch reference on 100% of 2,383 questions per checkpoint.

Set TYPESAFE_BASE_URL=http://localhost:11435 and the official SDK works unchanged, because POST /v1/systemone, /v1/decisions and GET /v1/models are wire-identical to TypeSafe.

Yes, with a Modelfile that sets FROM laya and QUESTIONS ./triage.json, plus parameters such as precision fp32. Then run ollaya create triage -f Modelfile and call it with ollaya run triage and the state.

It publishes only small ONNX graphs of about 3 MB and reads the original weight files from the author's Hugging Face repository pinned to a commit and verified by sha256, never re-hosting weights. GGUF models run on llama.cpp, the code is Apache-2.0, and each model keeps its own license.