d1
A decision model introduced by Liquid AI, belonging to the System One category, which leads on the Hugging Face Decision Index 0.2.1.
Core Facts
d1's decision calls appear in AI Gateway logs and are counted in budget statistics, allowing reuse of AI Gateway's cost attribution, call auditing, and quota control processes.
Liquid AI has expanded its D1 decision model with visual capabilities, enabling it to accept images, text, or both, and return probabilities for yes/no judgments, multiple-choice, or scoring questions.
Timeline (last 90 days)
There is currently no public benchmark data showing the performance of LFM3 and d1 on MLX.
At the time the article was published, the original release information did not disclose the specific parameter scales, architecture details or performance data of LFM3 and d1.
The commit message for this version contains 'mlx: add LFM3 and d1 models (#19333)', meaning the core change is adding the LFM3 and d1 models to the MLX backend.
Starting from TypeSafe's release of Jev in mid-September 2026, at least eight companies launched AI decision models within a little over a month, including Cloudflare's Clef, the AWS Strands team's Decider, Perplexity's pplx-decider, Liquid AI's d1 series, OpenAI's Decisions API and Microsoft-Decision-1.
Liquid AI's d1 scored 58.9 on the Hugging Face Decision Index 0.2.1, one point higher than Jev's 57.9.
On AI Gateway, d1 supports three access paths: the AI SDK decision API (experimental_decide/decide), the OpenAI-compatible Decisions API, and the TypeSafe-compatible API.
The AI SDK decision API is called through experimental_decide (alias decide), and requires passing three fields: model, state, and questions.
d1's decision calls appear in AI Gateway logs and are counted in budget statistics, allowing reuse of AI Gateway's cost attribution, call auditing, and quota control processes.
The combination of structured answers plus probabilities allows the caller to do threshold control: when the probability is below a certain value, the task is handed off to a human, and when it is above the threshold, it is executed automatically.
Whether decision models can stably surpass a "small model + prompt engineering" approach in accuracy remains to be verified, and the original text gives no benchmark data.
8 more timeline events
All Facts (14)
d1's decision calls appear in AI Gateway logs and are counted in budget statistics, allowing reuse of AI Gateway's cost attribution, call auditing, and quota control processes.
65%VerifiedLiquid AI has expanded its D1 decision model with visual capabilities, enabling it to accept images, text, or both, and return probabilities for yes/no judgments, multiple-choice, or scoring questions.
70%UnverifiedThere is currently no public benchmark data showing the performance of LFM3 and d1 on MLX.
50%UnverifiedAt the time the article was published, the original release information did not disclose the specific parameter scales, architecture details or performance data of LFM3 and d1.
50%UnverifiedThe commit message for this version contains 'mlx: add LFM3 and d1 models (#19333)', meaning the core change is adding the LFM3 and d1 models to the MLX backend.
50%UnverifiedThe AI SDK decision API is called through experimental_decide (alias decide), and requires passing three fields: model, state, and questions.
50%Unverifiedd1 supports three question types: Boolean, Choice, and Score, with answers accompanied by probabilities.
50%Unverifiedd1 represents a division-of-labor approach: generation tasks are given to language models, judgment tasks are given to decision models, and routing, classification, and scoring scenarios require low-latency, interpretable, verifiable output.
50%UnverifiedThe combination of structured answers plus probabilities allows the caller to do threshold control: when the probability is below a certain value, the task is handed off to a human, and when it is above the threshold, it is executed automatically.
50%UnverifiedIn visual classification scenarios, previously a separate classification model had to be trained, but with d1 a unified decision interface can be used to accomplish this.
50%Unverifiedd1's invocation paradigm is: receive shared state (state), give answers to predefined typed questions, and no natural language output is involved.
50%UnverifiedOn AI Gateway, d1 supports three access paths: the AI SDK decision API (experimental_decide/decide), the OpenAI-compatible Decisions API, and the TypeSafe-compatible API.
50%Unverifiedd1 supports visual input: images can be passed in as part of the state, and typed questions can be asked about image content, for visual classification, inspection, and scoring.
50%UnverifiedWhether decision models can stably surpass a "small model + prompt engineering" approach in accuracy remains to be verified, and the original text gives no benchmark data.
50%