Published
Updated
Ollaya serves local decision models on port 11435, and prints winnow:e4b at 0.722
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
ollaya-dev/ollaya, Apache-2.0, calls itself Ollama for decision models and says it is independent of Ollama and of TypeSafe. The README prints winnow:e4b at 0.722 against TypeSafe Jev at 0.738, and 89 ms for five questions on an RTX 4090. The item count behind that pair is not defined on the page we read. We did not install it.
恒星sun posted Ollaya on September 30. The repository is ollaya-dev/ollaya, Apache-2.0, 155 commits on the page we opened. The README calls it “Ollama for decision models” and says the project is independent of Ollama and of TypeSafe. The README does not name a person. The GitHub org is ollaya-dev.
The server accepts POST /v1/systemone and /v1/decisions on port 11435. The printed client setting is TYPESAFE_BASE_URL=http://localhost:11435. Weights stay on the authors’ Hugging Face repositories, pinned by sha256. Ollaya publishes ONNX graphs of about 3 MB. GGUF checkpoints, named in the README as winnow and jevk5, run through llama.cpp.
The README prints a comparison whose item count, under the label “typed-decisions,” is not defined on the page we read. winnow:e4b is 0.722 against TypeSafe Jev at 0.738, and the timing line is 89 ms for five questions on an RTX 4090. decider:4b is 0.680. The default 2B is 0.591. kev:4b is 0.669. kev:9b is 0.722. winnow, the 12B, is 0.702. clm is 0.357. laya:en is 8 to 10 ms for five questions on an RTX 4090. The README says fp32 exports match PyTorch on 100% of 2,383 questions per checkpoint.
The model names on that page are laya, laya:en, laya:multilingual, laya:typed-decisions, decider at 2B, 4B, and 0.8B, decider:2b-vision, kev at 0.8b, 4b, and 9b, decision and Decision 1.0 Eos, qwen3guard, nli, gliclass, von, winnow, clm, and jevk5. We did not open the search page on ollaya.dev, so the list stops where the README stops. Upstream licenses in the README are mostly Apache-2.0. The README marks nli:deberta-v3-large as MIT and llama.cpp as MIT.
Winnow’s 73.2 on JevBench is a different figure from 0.702 and from 0.722. jeff and laya-mlx are other local servers on this desk. Ollaya is a process on port 11435. We did not install it.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Post by @vintcessun, display name 恒星sun, September 30, 2026, 07:54 UTC. The first photo is the share image on this page. The repository is github.com/ollaya-dev/ollaya, Apache-2.0, 155 commits on the page we opened. The README says "Ollama for decision models" and that the project is independent, not affiliated with Ollama or TypeSafe. The README does not name a person. The GitHub org is ollaya-dev.
The server speaks POST /v1/systemone and /v1/decisions, on port 11435. TYPESAFE_BASE_URL=http://localhost:11435 is the printed setting. Weights stay on the authors' Hugging Face repos, pinned by sha256. Ollaya publishes ONNX graphs of about 3 MB. GGUF checkpoints, including winnow and jevk5, go through llama.cpp.
Printed comparisons, in the README's order. The item count of the "typed-decisions" label is not defined on the page we read. winnow:e4b 0.722 against TypeSafe Jev 0.738, and 89 ms for five questions on an RTX 4090. decider:4b 0.680. The default 2B is 0.591. kev:4b 0.669. kev:9b 0.722. winnow, the 12B, 0.702. clm 0.357. laya:en is 8 to 10 ms for five questions on an RTX 4090. fp32 exports match PyTorch on 100% of 2,383 questions per checkpoint.
The model names on that README are laya, laya:en, laya:multilingual, laya:typed-decisions, decider 2B, 4B, and 0.8B, decider:2b-vision, kev 0.8b, 4b, and 9b, decision and Decision 1.0 Eos, qwen3guard, nli, gliclass, von, winnow, clm, and jevk5. We did not open ollaya.dev/search, so this list is the README. Upstream licenses in the README are mostly Apache-2.0. nli:deberta-v3-large is MIT. llama.cpp is MIT. The site linked from the repo is https://ollaya.dev. We did not install Ollaya.
Compare
Winnow-12B Q8 at 73.2 on JevBench v1.5.4 is a different number from the README's winnow 0.702 and from winnow:e4b at 0.722. jeff and laya-mlx are other local servers already on this desk. jeff trails Jev on 1,600 public items. laya-mlx reports agreement with upstream Laya on 63 questions and does not add a JevBench row.
Ollaya listens on localhost:11435. It is not a public hosted door, and this desk does not list it next to TypeSafe, Vercel, or OpenRouter.
Terms
- 0.722
- winnow:e4b in the Ollaya README, against TypeSafe Jev at 0.738. The page also prints 89 ms for five questions on an RTX 4090. The item count for that pair is not defined on the page we read.
- 11435
- The port in the Ollaya README. POST /v1/systemone and /v1/decisions. TYPESAFE_BASE_URL is set to http://localhost:11435.
- 2,383
- Questions per checkpoint on which the README says fp32 exports match PyTorch, 100% of them.