Ollaya: Run Jev Decision Models Locally for Free

Ollaya, an Apache-2.0 open-source runtime that runs decision models locally, hit the Hacker News front page with nearly 500 points and offers a TypeSafe-compatible API as an alternative to TypeSafe's Jev, which launched eleven days ago with a $40 million seed round and a waitlist. Ollaya's Laya model (421M parameters) runs at 8–10 ms latency on an RTX 4090 versus Jev's 236–276 ms, at no cost versus Jev's $0.042 per million tokens, and posts a better calibration score (ECE 0.081 vs Jev's 0.246), though Laya's accuracy degrades on choices with more than 20 options. Decider-4b from Mapika, also available through Ollaya, ranked first on JevBench as of September 24, 0.8 points ahead of Jev.

TypeSafe’s Jev launched eleven days ago, pulled a $40 million seed round, and promptly ran into a waitlist. The pitch is genuinely interesting: a class of AI model that answers structured questions — route this ticket, score this urgency, flag this for spam — in milliseconds, without generating a single token of text. No hallucination, no output cost, answers your code can actually use. The only catch: you’re waiting in line. Ollaya https://ollaya.dev/ is the answer you didn’t need the line for. It’s an open-source runtime that runs the same category of models — decision models — on your own hardware. Apache-2.0 licensed, one binary to install, TypeSafe-compatible API out of the box. It hit Hacker News front page https://news.ycombinator.com/item?id=49848269 this morning with nearly 500 points. If you’ve been watching the Jev coverage and wondering whether there’s a self-hosted path, this is it. What Decision Models Actually Do If you’ve read ByteIota’s previous coverage of TypeSafe Jev, you know the category. If not, here’s the short version: decision models accept two inputs — a state any text or JSON and a set of typed questions — and return calibrated probabilities in a single forward pass. That’s it. No text generation, no token-by-token sampling. You ask “is this spam?” and get a probability. You ask “which category fits?” and get ranked options with confidence scores. This makes them genuinely useful for the judgment calls your code makes thousands of times a day. Routing support tickets. Scoring message urgency. Flagging tool calls before an AI agent executes them. Tasks where an LLM call is overkill and a regex is insufficient. How Ollaya Works Installation is a single command: curl -fsSL https://ollaya.dev/install.sh | sh The installer auto-detects an NVIDIA GPU driver R580+ and adds CUDA support automatically. On Windows, run the PowerShell equivalent from ollaya.dev/download https://ollaya.dev/download . Docker is also supported for server deployments. From there, you pull and run a model the same way you would with Ollama: ollaya run laya "Is this a spam message?" ollaya serve starts the TypeSafe-compatible API server Ollaya ships with several models. Laya 421M parameters is the fastest option and handles 100+ languages — it’s the right starting point for most classification tasks. Decider 2B parameters, Qwen3.5-based trades speed for accuracy. NLI and GLiClass handle zero-shot and named-entity classification. Each model ships as a verified ONNX graph; you’re not trusting an unreviewed fine-tune. The Numbers That Matter Local decision models are faster and cheaper than hosted. That’s expected. What might surprise you: the accuracy gap is smaller than the price gap, and on calibration — the reliability of confidence scores — open models actually lead. | | Jev Hosted | Ollaya + Laya Local | |---|---|---| | Cost | $0.042 per million tokens | Free | | Latency | 236–276 ms | 8–10 ms RTX 4090 | | Calibration ECE | 0.246 | 0.081 better | | Choice 20 options | Handles well | Accuracy degrades | | Data privacy | Leaves your machine | Stays local | Laya’s calibration score ECE 0.081 vs Jev’s 0.246 is worth pausing on. Better calibration means confidence scores you can actually threshold on. When you’re deciding whether to escalate to a human at 80% confidence, you need the model to mean what it says when it says 80%. Laya does that better than Jev. Decider-4b from Mapika — also available through Ollaya — is currently ranked first on JevBench https://laya-ai.com/laya-vs-jev as of September 24, sitting 0.8 points ahead of Jev. These are not toy alternatives. One Line to Switch Ollaya implements the TypeSafe API endpoints, so your existing SDK code requires one change: python from typesafe import Client Before: base url="https://api.typesafe.ai" client = Client base url="http://localhost:11434", api key="local" Your question schemas, response handling, and threshold logic stay exactly the same. This makes it straightforward to run Ollaya in development and CI, then evaluate at deployment time whether local accuracy is sufficient or you need the hosted API for edge cases. When Local Is Enough For most routine classification work — routing tickets, scoring urgency, filtering spam, screening content — the open models perform well. Where they break down: choice questions with more than 20 options, and tasks requiring multi-hop reasoning or precise date comparisons. Jev has the same limitations on reasoning; it’s just more polished on edge cases in classification. The pattern emerging in 2026 agentic systems is three layers: deterministic code for fully specified logic, decision models for bounded judgment calls, LLMs for open-ended reasoning and generation. Ollaya fits cleanly into the second layer. Use it for the thousands of classification calls that don’t need frontier accuracy. Reserve the LLM calls for the ones that do. The Open Model Race It took eleven days for a front-page Hacker News thread on a fully open-source local alternative to TypeSafe’s funded product. This is the LocalStack/Terraform pattern repeating: VC-backed tool ships, community ships the open equivalent, the two coexist for different use cases. TypeSafe’s real moat is its proprietary RLCD training pipeline and the accuracy it produces on the hardest classification tasks. But for the 80% of calls that don’t live at that edge, Ollaya is already competitive and open https://github.com/ollaya-dev/ollaya . Run curl -fsSL https://ollaya.dev/install.sh | sh and test it against your own use case before committing to the API bill.