ollaya
Ollama for decision models: pull laya kev von decider, serve POST /v1/systemone on :11435, ONNX graphs only, TypeSafe SDK drop-in.
Overview
ollaya (github.com/ollaya-dev/ollaya, Apache-2.0, ollaya.dev) is a Rust CLI and daemon that downloads small ONNX graphs (about 3 MB each), verifies upstream Hugging Face weight blobs by sha256, and serves choice, score, and noul questions over POST /v1/systemone, /v1/decisions, and GET /v1/models. curl install.sh, then ollaya run laya --preset triage on a ticket string prints intent bars in milliseconds. TYPESAFE_BASE_URL=http://localhost:11435 works with the official SDK unchanged. Registry includes laya routers, decider, kev, von, nli, gliclass, and qwen3guard without re-hosting author weights. ollaya mcp and the ollaya-decisions skill target agent clients. Modelfiles bake question sets into custom tags.
Problem: Pulling each open decision checkpoint by hand means different serve flags, ONNX exports, and SDK base URLs for Laya, Kev, Von, and friends.
Built for: Developers who want Ollama-like pull and run for decision models with TypeSafe SDK compatibility on localhost:11435.
First indexed on Jev Directory: 2026-09-25
Creator and team
- Name
- ollaya
- Handle
- @ollaya-dev
How Jev is used
- Role in the product flow
- Local ONNX decision runtime exposing TypeSafe wire-identical System One HTTP
- Primitives
- ChoiceScoreNoul
- State in
- JSON state plus typed questions; CLI presets and Modelfile QUESTIONS JSON for bundled triage sets per README.
- Decision out
- Calibrated probability bars per question in one forward pass without text generation.
- Install via ollaya.dev/install.sh and start ollaya serve on port 11435
- ollaya pull kev or ollaya run laya --preset triage on sample text
- Point TypeSafe clients at TYPESAFE_BASE_URL=http://localhost:11435
- Optional ollaya mcp for Claude Code, Cursor, and Desktop agent tooling
ollaya is infrastructure, not another Laya marketing page: one daemon serves many open decision checkpoints with the same HTTP shape. Bud Decision Studio (budecosystem-bud-decision-studio) is the GUI desktop alternative: Playground, model downloads, and localhost :8420 for the same SDK contract. jaredpalmer-kev and wfzyx-von remain the authoritative repos for those models; ollaya pulls verified ONNX wrappers. togethercomputer-tev1 is a training recipe on Together; ollaya is local pull and serve. githubnext-localjev bridges chat JSON on Mac; ollaya runs native decision graphs. Do not list Laya itself again here; use ollaya when the story is runtime packaging.
Sourced performance claims
- README documents laya:en at 8 to 10 ms for five questions on RTX 4090 and fp32 parity with PyTorch on 100% of 2,383 questions per checkpoint.Source: github.com/ollaya-dev/ollaya README Models and Features
- decider:0.8b listed at 0.591 accuracy on typed-decisions in the model table.Source: github.com/ollaya-dev/ollaya README Models table
- Public GitHub repo had about 153 stars when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- Ollama-style pull, list, ps, run, and Modelfile create workflow
- TypeSafe-compatible /v1/systemone on default port 11435
- Weights stay on author HF repos; ollaya ships pinned ONNX graphs only
- MCP server and ollaya-decisions agent skill
- Desktop app and Docker cuda images on ollaya.dev/download
Stack
- Rust
- ONNX Runtime
- CUDA and CPU
- TypeSafe HTTP contract
Pricing: Apache-2.0 binary; model licenses follow each upstream author listed in README.
Links
FAQ
- Does ollaya replace downloading Kev or Von directly?
- It wraps upstream weights with ONNX graphs and a unified CLI. Author repos remain the source of truth for training details.
- Will this create a duplicate Laya tools page?
- This listing covers the ollaya runtime only. It does not re-publish Convai Laya as a separate product profile.
- Which port should SDK users set?
- README documents TYPESAFE_BASE_URL=http://localhost:11435 for wire-identical System One calls.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.