Loading the blog…
WRITING
Articles on System One decision models — how they differ from LLMs, the Laya and Jev architectures, calibration, and publishing to the registry.
We're excited to share that System One Models is a member of NVIDIA Inception, the NVIDIA Inception program for technology startups.
Read the articleHow to put your certificate of completion on your LinkedIn profile in one click, and how a recruiter or a colleague checks that it is real: the certificate's page, its number and its QR code.
Getting started with System One models is a free course in nine short lessons, from what a decision model is to publishing your own. Finish it and claim a certificate with your name that anyone can check.
Twenty decision models and APIs joined the registry this week: OpenAI's Decisions API, Liquid AI's d1, Upstage's Solar Decide, two more hosted APIs, and fifteen open models from InternLM, Juspay and others.
The open standard for System One models has a new name: NoulXP. Same format, same answers, same badge. Install it with pip install noulxp; everything made with OpenDXP still runs.
We served Julia 1, Laya, Decider 2B and our package of AnyJev L0 (Nokia's method) on Qwen3-1.7B on one rented NVIDIA A40 and measured decisions per second, latency, cost per million decisions and how many models one GPU can hold. The cheapest cost $0.46 per million decisions, the most expensive $11.79.
torch.onnx.export(dynamo=True) with named torch.export.Dim objects fails its first two capture strategies on torch 2.8 and falls back to draft_export, four times slower. Naming the dimensions as strings fixes it.
We moved six decision-model packages from an Apple M4 to an NVIDIA A40 and an Intel Xeon, and checked every probability against answers recorded on the Mac. Several things changed the answers, none of them raised an error, and two settings fixed most of it.
We quantized our package of AnyJev L0 (Nokia's method) on Qwen3-1.7B to Q8_0 and Q4_K_M and checked every answer against the model's own. On an NVIDIA A40, 4-bit flipped 14 of 90 decisions, 8-bit flipped 2, and neither was faster.
Asking a language model a multiple-choice question once per rotation of the options removes its position bias. On an NVIDIA A40, a prompt cache cut the cost of those rotations by 2x, decoding only a third of the tokens, with every decision unchanged.
Six decision models on one NVIDIA A40: two causal models' prompts made an 8 ms model wait 129 ms. A thread per model, with four computing at once, raised the total from 45 to 80 decisions per second at 16 clients per model, and from 44 to 69 at one.
Agents make many small decisions, and asking an LLM for each one is slow and gives no usable confidence. Here's how to give Claude, Cursor or any MCP client an open decision model as a tool that answers in a fraction of a second, with a probability for every option.
Most decision models we tested were surer than they were right. Three numbers per model, fitted to 250 labels, brought one from 96% confidence down to its real 72% without touching its weights. What you label with, and which data you use, matter as much as how many.
Batching is supposed to be free throughput. For decision models it slowed an Apple M4's CPU by up to 2.2x. On an NVIDIA A40 it paid off once each pass held only rows of similar length, and then the same GPU served 1.6x more.
A model that returns an answer from a set you define, with a calibrated probability, instead of writing text. How it works, how it differs from an LLM, a classifier and JSON mode, and when to use one.
Jev's weights are not published. Within ten days of its release, more than twenty open System One models appeared that you can run and fine-tune yourself, and eleven of them answer Jev's own API. What each one is, and how to pick.
A general model hub can store a decision model perfectly well. Storage was never the problem — the metadata, the evaluation and the workflow are.
A System One model returns a decision, not a paragraph. That single difference changes how you evaluate it, deploy it and reason about it.
NoulXP 0.2 adds an HTTP API, an MCP binding for AI agents, and rotations. The first model asked the new way, Nokia's AnyJev on Qwen3-1.7B, makes the same decision as Nokia's own code on all 90 test questions.
Tev1-4B-experimental is a LoRA fine-tune of Qwen3.5-4B that reads a state and up to 24 options and answers with one letter. Together published the model, the recipe and the tutorial; the weights licence is still being decided.
LLMs and decision models are not competitors. They answer different questions, and using the wrong one is expensive in a way that is easy to miss.
A walk through the whole loop — write a manifest, create a repository, publish a version, record the lineage — in about ten minutes.
System One Studio now trains on PyTorch as well as MLX, so it runs on Windows and Linux too. We ran the same fine-tune on both on a 16 GB M4 — time, memory, accuracy, calibration, and whether the two checkpoints agree — and found a bug on the way.
Laya is an open-weights System One model built on a ModernBERT encoder. Here is how it scores options in one pass, what its checkpoints are for, and how to fine-tune it on a laptop.
Within a week of Jev, individuals and small teams had rebuilt its shape from open bases. Some publish better calibration than the original. A guide to the ones worth knowing.
Jev is the model that named the category. What it takes in, what it returns, what it costs, and what its makers have not said.
Kev is an open family of System One models from Jared Palmer that keeps a Qwen base frozen and learns only an adapter and a pointer head. Small to download, honest about calibration, and it speaks Jev's API.
Yes. Hugging Face hosts every kind of model; System One Models (systemonemodels.ai) is a registry built only for System One models, with the facts a decision model needs. What each does, side by side, and how to use them together.
Fastino's open-weight decision model does something the others do not. It classifies against label sets you hand it per request, extracts the spans that justify the answer, and enforces rules across decisions, all in one pass of a 340M encoder.
Fifteen model families have appeared in ten days, from a closed hosted API to a 2.8 MB form-filler. What each one decides, what it costs to run, and what its makers actually measured.