Macrocosmos’ First Public Model, IOTA Decide 4B, Helps Systems Make Fast, Structured Decisions

A customer support ticket arrives, and software must decide where it belongs, how urgent it is, and whether it needs human attention.

Macrocosmos has released IOTA Decide 4B to handle this kind of work. The compact model returns probabilities for predefined options, helping applications make clear, structured decisions.

IOTA Decide 4B is Macrocosmos’s first public model. Its weights are available on Hugging Face under the Apache 2.0 license, allowing developers to download, adapt, and deploy it. The release also demonstrates the training infrastructure behind IOTA (SN9).

The model accepts information such as a support ticket, an AI agent’s activity log, or a policy paired with a customer request. Developers supply questions with possible answers, including yes or no choices, categories, and severity levels. The model returns a probability for every option in a single processing pass.

Applications can use these results to route requests, check eligibility, flag risky activity, and apply guardrails to AI agents. Confidence scores also help determine which cases need human review. This gives developers a direct way to connect model outputs to actions within their software.

Macrocosmos reports a JevBench composite score of 78.6, placing IOTA Decide 4B in the top 10 among more than 300 open and closed systems. The benchmark evaluates capability, calibration, speed, intelligence, and cost. Despite having just four billion parameters, the model finishes within roughly three points of the leader.

Its reported scores include 92.2 for speed, 87.8 for calibration, 83.0 for cost efficiency, 74.0 for capability, and 60.2 for intelligence.

Macrocosmos also reports processing times as low as 33 milliseconds per decision on a single GPU. Estimated operating costs are about $0.0037 per 1,000 decisions when developers host the model themselves.

The training process is another central part of the release. IOTA Decide 4B is built on the Qwen3.5 4B foundation. It trained the model on approximately 879,000 decision examples using the IOTA SDK. These combined examples were labeled by humans, public datasets, and synthetic cases.

Training ran across 16 geographically distributed RTX A4000 GPUs connected over the internet. This was coordinated using streaming DiLoCo (Distributed Low-Communication), which allows machines in different locations to train a shared model. Macrocosmos’s distributed batch inference infrastructure also supported data generation and labeling.

IOTA Decide 4B gives Macrocosmos a concrete demonstration of its Liquid Compute approach. It connects the infrastructure on SN9 to a practical product for developers who need fast, affordable decision processing. Macrocosmos describes it as the first of more open models to come.