IOTA Decide 4B

IOTA Decide 4B is Macrocosmos’s first public model, announced 2026-11-04: a ~4B decision model (not a chat LLM). Given a state (ticket, agent trace, policy+request, etc.) plus typed questions, it returns answers with calibrated probabilities in a single forward pass—no free-form generation or prompt crafting.

Built from Qwen3.5-4B with Apache-2.0 weights at macrocosm-os/iota-decide-4b (plus GGUF). Training is framed around distributed GPUs on Bittensor SN9 / IOTA. The API follows SystemOne /v1/systemone. Macrocosmos self-reports ~78–79% on the public JevBench split.

Problem

Many agent/business loops need structured decisions with confidence—not essays. Decide 4B ships a joint multi-question decision head.

Differentiation

Versus chatty 4B LMs: labels/scores/option probs. Versus closed decision APIs: self-hostable Apache weights. Tied to Bittensor SN9/IOTA training narrative.

Limits

• Still early HF traction and third-party evals.
• Sealed JevBench official scores may still be pending.
• “IOTA” is overloaded—disambiguate with Macrocosmos/Decide.

Who it’s for

Teams needing batch structured decisions (routing, policy checks, trace judgments) with self-hosted weights.

Get started

1. Open the HF card.
2. Call systemone(...) or serve the GGUF with llama.cpp’s /v1/systemone.
3. Calibrate probability thresholds on your own tickets/policies before production.