# Ollaya

> System 1 for AI agents. Ollaya serves open decision models on your own machine, behind a self-hosted, Jev-compatible API: typed, calibrated answers in milliseconds for routing, guardrails, triage and tool approval.

A decision model reads a state (a message, an email, a ticket, any JSON) plus typed questions (choice, score, yes/no) and returns calibrated probabilities in a single forward pass, in milliseconds. It never generates text. Ollaya pulls these models by name from their authors' Hugging Face repositories and serves them from a local daemon on port 11435 behind a self-hosted, Jev-compatible API (TypeSafe's `/v1/systemone` wire format), so the TypeSafe SDK and other Jev clients work unchanged once they point at Ollaya and name an open model. Install: `curl -fsSL https://ollaya.dev/install.sh | sh`.

## Docs
- [Quickstart](https://ollaya.dev/docs/quickstart): Install Ollaya, run your first decision model and call the local API.
- [From the browser to your code](https://ollaya.dev/docs/browser-to-code): Take the model you tried in your browser to your own code. Pull coba:150m-ternary with Ollaya and call it from curl, Python, TypeScript or a Docker container.
- [CLI reference](https://ollaya.dev/docs/cli): Every ollaya command and flag: run, pull, serve, list, ps, show, stop, rm, cp, create, preset, update and mcp.
- [API reference](https://ollaya.dev/docs/api): The native /api endpoints, the TypeSafe-compatible /v1 endpoints, errors, streaming and limits.
- [Modelfile](https://ollaya.dev/docs/modelfile): Bake a question set, calibration, precision and license into a model you can run by name.
- [TypeSafe compatibility (Jev-compatible API)](https://ollaya.dev/docs/typesafe-compatibility): Ollaya is a self-hosted, Jev-compatible API server. Use the official TypeSafe SDK and existing Jev code with local, open models by changing environment variables.
- [Open-source Jev alternative: which open model on which hardware](https://ollaya.dev/docs/jev-alternative): Replace TypeSafe's hosted Jev with an open decision model on your own machine, behind the same API. Which model to run on a 4, 8 or 12 GB GPU or a CPU, with measured accuracy, calibration and speed.
- [Agents: MCP and skill](https://ollaya.dev/docs/agents): Give AI agents local, typed decisions: the Ollaya MCP server for Claude Code, Claude Desktop, Cursor and other clients, and an agent skill that teaches when and how to use it.
- [Guardrails: prompt injection and jailbreak detection](https://ollaya.dev/docs/guardrails): Check prompts, tool results and retrieved documents for prompt injection, jailbreaks and harmful content on your own machine, with open guard models and calibrated probabilities.
- [Routing: pick the model for each request](https://ollaya.dev/docs/routing): Route each request to the right LLM, tool or team with a local decision model. Difficulty, domain, tool use and sensitivity come back typed and calibrated, in milliseconds.
- [Replacing LLM classification calls with a local model](https://ollaya.dev/docs/replace-llm-classifiers): Move the classify, triage and yes/no prompts you send to a hosted LLM onto a local decision model, with typed answers, calibrated probabilities and no per-token cost. Nothing leaves your machine.
- [Zero-shot classification on a CPU](https://ollaya.dev/docs/cpu-classification): Classify text into labels you choose per request, on a CPU with no GPU, in under a second. Which open models to run, and how fast and accurate they are, measured.
- [FAQ](https://ollaya.dev/docs/faq): How Ollaya relates to Ollama and TypeSafe, hardware, privacy, where weights come from, and licensing.
- [For model authors: Day-0 support](https://ollaya.dev/docs/model-authors): Releasing a decision model? Share it with Ollaya before launch and `ollaya run` works on announcement day, with parity against your code and benchmark numbers you can quote.
- [Integrations: LiteLLM, LangChain, AI SDK, DSPy, LlamaIndex, n8n and Dify](https://ollaya.dev/docs/integrations): Use Ollaya from LiteLLM, LangChain, the Vercel AI SDK, DSPy and LlamaIndex through their TypeSafe integrations, with a base URL and nothing else to install, and from n8n and Dify workflows over HTTP.

## Models
- [winnow](https://ollaya.dev/library/winnow): Decision models by EldanRing, fine-tuned from Google's Gemma 4 and published as GGUF. Winnow reads the answer labels' logits after its own prompt; Ollaya runs the author's file on llama.cpp, on NVIDIA GPUs, Vulkan GPUs on Windows, Apple silicon or the CPU.
- [laya](https://ollaya.dev/library/laya): Open decision models from Convai Innovations. Typed, calibrated answers to choice, score and yes/no questions in a single forward pass, in English and 100+ languages.
- [decider](https://ollaya.dev/library/decider): Decoder decision models by Mapika on Qwen3.5: the answer is read from option-letter logits in one forward pass.
- [kev](https://ollaya.dev/library/kev): Decision models by Jared Palmer: a LoRA on a Qwen3.5 base plus a pointer head that scores every option at its own span, in one forward pass per question. Calibrated with Kev's own temperature.
- [nli](https://ollaya.dev/library/nli): Zero-shot classifiers by Moritz Laurer: every option becomes a hypothesis scored for entailment. The most accurate encoder model on typed decisions in our tests.
- [gliclass](https://ollaya.dev/library/gliclass): Instruction-following zero-shot classifiers by Knowledgator: all options of a question are scored in one pass, so on a GPU the cost barely grows with the number of options; on a CPU it grows with them.
- [qwen3guard](https://ollaya.dev/library/qwen3guard): Safety guard by the Qwen team: is a text safe, controversial or unsafe, and which unsafe category? It answers its own built-in questions, in 119 languages, in one forward pass.
- [d1-omni](https://ollaya.dev/library/d1-omni): Liquid AI's d1-omni-600M, text questions: an LFM2.5 encoder with a decision head that scores every option at its own marker, all of a question's options in one pass, calibrated per question type and option count. 16k-token context. The checkpoint's image and audio paths are not served.
- [decision](https://ollaya.dev/library/decision): Decision models by the vLLM Semantic Router contributors: a fully fine-tuned Qwen3.5 backbone plus an endpoint head that scores every option at its own last token against the question, in one forward pass per question. 16k-token rows.
- [von](https://ollaya.dev/library/von): Decision model by Victor Hugo Panisa on ModernBERT-large: every option is scored at its own marker, all options of a question in one pass, with an input-conditioned calibration. 8k-token context.
- [jevk5](https://ollaya.dev/library/jevk5): Decision model by alibiserikbay, fine-tuned from Qwen3.5-4B and published as GGUF. JevK5 reads the answer letters' logits after its own JSON prompt; Ollaya runs the author's file on llama.cpp, on NVIDIA GPUs, Vulkan GPUs on Windows, Apple silicon or the CPU.
- [clm](https://ollaya.dev/library/clm): Contrastive decision model by Contrastive-LM: the Qwen3-8B encoder embeds the state and every option, and two trained heads pick the option closest to the state. Repeated questions and options are cached.
- [gliclass-v4](https://ollaya.dev/library/gliclass-v4): Knowledgator's GLiClass v4, the decoder-KV generation: a Qwen3.5 decoder (2B or 9B) reads the instructions, the state and the labels in one row, a two-layer scorer with up to three reasoning steps scores every label against the text, and the model's own calibrator rescales the logits. Up to 8k-token rows. Text only.
- [nimble](https://ollaya.dev/library/nimble): Decision model by Bespoke Labs: a LoRA on Qwen3.5-9B trained on contrastive pairs. Nimble reads the whole request as a JSON schema and scores each option by the next-token logit of its code; Ollaya applies the author's temperature, so its probabilities are calibrated. Up to 255 options.
- [jeb](https://ollaya.dev/library/jeb): Decision models by Jason Brashear and AINode: LoRAs merged into Qwen3.5 and Qwen3.8, published as GGUF. Jeb reads the option labels' logits after AINode's own decision prompt, with a fitted temperature per question type; Ollaya runs the authors' files on llama.cpp.
- [cygnet](https://ollaya.dev/library/cygnet): Frozen Gemma 4 12B IT with blockbrain-ai's Cygnet prompt: the options as letters, one answer slot, and one calibration temperature. No fine-tuning; Ollaya runs ggml-org's Q8_0 GGUF on llama.cpp.
- [jeeves](https://ollaya.dev/library/jeeves): Decision model by PostHog: Qwen3.5-9B with a LoRA merged in and a pointer head that scores every option at its own marker. Ollaya runs it without its reasoning chain, in one forward pass per question, calibrated with the authors' temperature.
- [clef](https://ollaya.dev/library/clef): Decision models by Cloudflare. Clef-Flash is Qwen3.5-9B, fully post-trained, with a joint schema head that scores every option of every question together, in one forward pass per request. Its probabilities come straight from the head. Ollaya runs its text path.
- [snap](https://ollaya.dev/library/snap): logitlab's snap1-2b: MiniCPM5-2B fine-tuned to read one option letter after the prompt of emnlmn's snap engine. Ollaya runs the author's Q8_0 GGUF on llama.cpp with snap's own prompt, on NVIDIA GPUs, Apple silicon or the CPU.
- [decima](https://ollaya.dev/library/decima): A. M. Madani's Decima: multilingual encoders (mmBERT-base, multilingual-e5-small) with a late-interaction scorer that reads every option against the state, so option order never changes the answer, and an ordinal head for scores. A general model, one for coding-agent decisions, two guards for LLM apps, and a small one that is the fastest on a CPU.
- [arbiter](https://ollaya.dev/library/arbiter): Codekins' Arbiter (Zyot Lab): Gemma 3 4B IT with a LoRA and a fixed 24-slot head that reads every option at the end of one prompt per question. It answers yes/no questions, choices of up to 16 options and scores of exactly 6 levels.
- [credence](https://ollaya.dev/library/credence): Txoka's Credence v1: MiCA refinements of EldanRing's Winnow-E4B (Gemma 4), run from the author's Q8_0 GGUFs on llama.cpp. Two checkpoints trade public accuracy against calibration, each with a vision tag that adds Winnow-E4B's unchanged projector; neither beats original Winnow on every metric.
- [strands-decider](https://ollaya.dev/library/strands-decider): Strands Agents' decision model: a LoRA and a pointer head on Qwen3.5-2B base or on Gemma 4 E2B / E4B. Each question is one row with a line per option, every option scored at the end of its own line, and the probabilities are calibrated with the authors' temperature for each question type. Up to 255 options.
- [pplx-decider](https://ollaya.dev/library/pplx-decider): Perplexity's decision model: Qwen3.5-4B distilled from its 27B decision models, with a readout over 255 answer codes. Every question of a request shares one sequence and reads the whole request, and the probabilities are calibrated with Perplexity's temperature. Up to 255 options.
- [largha](https://ollaya.dev/library/largha): Phocinae's Largha: Laya's decision network on mmBERT-small (144M), fine-tuned on the typed-decisions workflows. Small enough for a CPU, on the Apple GPU through MLX.
- [coba](https://ollaya.dev/library/coba): Ollaya's own decision model: Laya's network on Ettin encoders, distilled from pplx-decider:flash. English. Full-precision 1b and 400m, and ternary sizes small enough for a browser.
- [sura](https://ollaya.dev/library/sura): Ollaya's own decoder decision model: Qwen3.5 distilled from pplx-decider:flash in its joint format. Every question of a request shares one sequence and reads the whole request; up to 255 options. English, zero-shot.
- [d3](https://ollaya.dev/library/d3): Decision 3.0 by the vLLM Semantic Router contributors: a fine-tuned Qwen3.5 whose full-attention layers read the whole prompt, with a readout over 255 answer codes. Every question is its own forward pass. Five sizes, 0.6B to 9B.
- [decision2](https://ollaya.dev/library/decision2): Decision 2.0 by the vLLM Semantic Router contributors: a fully fine-tuned Qwen3 or Qwen3.5 backbone plus an endpoint head that scores every option at its own last token against the question, in one forward pass per question. Five sizes, 0.6B to 9B.
- [d1](https://ollaya.dev/library/d1): LiquidAI d1-3B: typed decisions with max-pooled answer forms, Q8_0.

## Optional
- [Results](https://ollaya.dev/results): accuracy, calibration, speed and parity of every model, with Ollama on the same GPU
- [Ollaya Bench](https://bench.ollaya.dev): the public leaderboard of open decision models
- [Source code](https://github.com/ollaya-dev/ollaya): Apache-2.0
- [Full docs](https://ollaya.dev/llms-full.txt): every docs page above in one Markdown file