News · SI
Macrocosmos releases IOTA Decide 4B, an open model that answers with probabilities instead of text
Macrocosmos, the team behind Bittensor's IOTA subnet, has released IOTA Decide 4B, its first public model. Built on Qwen3.5-4B and fine-tuned across 16 internet-connected GPUs, it returns calibrated probabilities for routing, policy and triage decisions in a single pass, under Apache 2.0.
Macrocosmos, the company behind the IOTA subnet on the Bittensor network, has released IOTA Decide 4B, its first public model[3]. It's an open-weights "decision model": rather than writing a paragraph, it hands back calibrated probabilities that software can act on directly[1]. The weights are Apache 2.0[1].
01 / 04
Answers, not essays
You give IOTA Decide a state, which can be a support ticket, an agent trace, or a policy paired with a request, plus a set of typed questions about it[1]. Each question gets one of three kinds of answer: the probability a statement is true, a chosen option with probabilities for every option, or an expected score on a rubric[2]. That means a developer can route a ticket or decide when to send a case to a human with ordinary threshold logic, without parsing free text back into a category[2]. The interface follows the SystemOne /v1/systemone format, so existing decision-API clients work unchanged[1].
02 / 04
Trained without a GPU cluster
The model builds on Qwen3.5-4B and was trained on about 879,000 decision examples, using streaming DiLoCo across 16 RTX A4000 GPUs spread around the world[3]. Macrocosmos says the training data also came from its own stack, generated and labelled with distributed batch inference[3]. It's built with the company's IOTA SDK, which Macrocosmos says will be released publicly soon[1].
03 / 04
How it scores
TAO Media reports a JevBench composite of 78.6, with 87.8 for calibration and 92.2 for speed, and about 33 milliseconds per decision[2]. In Macrocosmos's own evaluation the model got 78.8% of the 231-question public JevBench split right, and 74.2% on domains it never saw in training[1]. The company is upfront about weak spots: multi-step arithmetic like date windows and pro-rated amounts, and sometimes accepting a response that misses a requirement. It advises keeping a human in the loop for high-stakes medical, legal, credit or hiring decisions[1].
04 / 04
Why it matters
Many production uses of language models are really classification and routing jobs dressed up as chat. A 4B model that runs on a single GPU or even a CPU, and returns probabilities you can threshold, is a cheaper and more predictable fit for that work. It's also a working example of a useful model fine-tuned on decentralised, internet-connected hardware, which is the core bet behind Bittensor's training subnets.
FAQ
Frequently asked questions.
01What does IOTA Decide 4B do?
You give it a state, such as a support ticket, an agent trace or a policy plus a request, along with typed questions. It returns a calibrated answer to each question, like a yes/no probability, a chosen option with confidence, or a rubric score, without generating any text.
02How was it trained?
It's a fine-tune of Qwen3.5-4B on about 879,000 decision examples, trained across 16 RTX A4000 GPUs connected over the internet using streaming DiLoCo rather than a single GPU cluster.
03How well does it score?
TAO Media reports a 78.6 JevBench composite and roughly 33 milliseconds per decision. Macrocosmos's own run on the 231-question public JevBench split got 78.8% accuracy, and 74.2% on domains never seen in training.
04Where can I get it and what's the license?
The weights are on Hugging Face under Apache 2.0. They run on a single GPU (about 9 GB in bf16) or even on CPU, and use the SystemOne /v1/systemone request format.
That’s the story
Building with what’s new?
Spectrum Web Co ships products, places talent, builds social narratives and builds custom business tools as monthly services.
Related stories