Skip to main content
Decisions need recent SDK versions: Python tinfoil >= 0.15.0, Go tinfoil-go >= v0.15.10, JavaScript tinfoil >= 1.3.0. Swift and Rust do not support decisions yet.
Decision models answer typed questions about an input and return calibrated probabilities instead of generated text. See the Decision Outputs guide for more.
Decision models are a rapidly changing area right now, hence the generic decision-small, decision-medium, and decision-large, reflecting their relative size tiers. Today all three are served by the vllm-sr d3 family, which perform well on the Jev Decision Index. As better decision models appear, the checkpoints behind these names may change.
Decision modelsDecision models
Decision Large
decision-large
Parameters: 27BQuantization: None (served in BF16)Capabilities: Typed decisions: true/false probabilities, choices from your options, and rubric scoresContext Window: 32k tokensBest for: Tricky decision tasks. The slowest of the family, though still quicker than outputting text.Model weights: vllm-sr/d3Configuration repo: tinfoilsh/confidential-decision-models
Decision modelsDecision models
Decision Medium
decision-medium
Parameters: 9BQuantization: None (served in BF16)Capabilities: Typed decisions: true/false probabilities, choices from your options, and rubric scoresContext Window: 32k tokensBest for: Most decision tasks: escalation, policy evaluation, triage. The default choice.Model weights: vllm-sr/d3-flashConfiguration repo: tinfoilsh/confidential-decision-models
Decision modelsDecision models
Decision Small
decision-small
Parameters: 4BQuantization: None (served in BF16)Capabilities: Typed decisions: true/false probabilities, choices from your options, and rubric scoresContext Window: 32k tokensBest for: Very fast, high-volume decisions, with some accuracy tradeoff. Works well as a first pass.Model weights: vllm-sr/d3-miniConfiguration repo: tinfoilsh/confidential-decision-models