Definition
Decision models perform structured classification and scoring over predefined option sets with calibrated confidence — faster and cheaper than free-form LLM generation for agent routing, guardrails, and triage.
Key Points
- 2026-10-09: microsoft-decision-1 (Qwen3.5-9B) on Foundry; typesafe-ai 7.5B for jev — capital + SKU race in System One layer (2026-10-09-microsoft-decision-1-model-foundry, 2026-10-09-typesafe-jev-870m-7-5b-valuation)
- 2026-10-07: AWS Strands Decider 2B — 2B params, pointer head on Qwen3.5-2B, Apache-2.0, ~115ms median latency (strands-decider-2b)
- Paradigm shift from token generation to single-forward-pass option scoring; complements system-one-models and jev architecture
- Fits between traditional classifiers (narrow, expensive to train) and LLMs (slow, overkill for routing)
Related
- strands-decider-2b
- system-one-models
- calibrated-decisions
- jev
- microsoft-decision-1
- typesafe-ai
- agent-routing
- agentic-ai
- llm-reliability