Nimble

Product Introduction

1. Definition: Nimble is a hosted API service providing specialized decision-making AI models. Technically, it falls under the category of "decision models" or "reasoning models," which are fine-tuned to output calibrated probabilities for structured choices rather than generating open-ended text.
2. Core Value Proposition: Nimble exists to turn ambiguous text into actionable, quantified decisions. Its core value is providing a probability score for every possible answer, enabling software to act autonomously on high-confidence calls and flag low-confidence decisions for human review. This bridges the gap between deterministic code and the probabilistic nature of real-world language.

Main Features

1. Multi-Question, Single-State Processing: A single API request can contain multiple decision questions (e.g., "Which department?" and "Is a refund requested?") based on one shared context ("state"). This architecture is efficient, as the model reads the context once to answer all related queries, reducing latency and cost compared to sequential calls.
2. Typed Answer Schemas (noul, choice, score): Nimble forces structured outputs through predefined answer types. The noul type returns a probability (0-1) for a binary true/false statement. The choice type selects from a user-defined list and provides a probability distribution across all options. The score type places an input on an ordinal scale (e.g., 1-5) and returns an expected value with per-level probabilities.
3. Model Fleet with Unified API: The service offers two distinct model sizes—Nimble-v4 (9B parameters) and Nimble-v4-max (27B parameters)—through a single, consistent API endpoint and request format. This allows developers to seamlessly switch between models optimized for cost/latency (v4) and maximum accuracy (v4-max) without code changes.

Problems Solved

1. Pain Point: The "last-mile" problem in AI automation: traditional language models generate text, but integrating their unstructured, often verbose outputs into deterministic business logic (like routing a ticket to a specific queue) requires complex, brittle parsing and validation logic.
2. Target Audience: The primary users are software engineers, ML engineers, and product managers building automated workflows. Specific personas include: SaaS developers implementing AI-powered customer support triage, content moderation platform engineers needing consistent policy enforcement, and data science teams operationalizing agentic AI systems that require reliable next-step decisions.
3. Use Cases: Essential scenarios include: automated customer support ticket routing and tagging, content moderation and policy compliance checking, grading and evaluation of structured responses, gatekeeping decisions for multi-step AI agents (e.g., "should the agent search the web?"), and dynamic form-filling from unstructured text.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):
   • Pros: Industry-leading benchmark performance (Decision Index), especially for model size; clean, probability-based outputs that integrate directly into code; cost-effective tiering with a smaller, faster model for high-volume tasks; hosted API eliminates infrastructure management.
   • Cons: As a hosted API, it introduces external dependency and potential latency vs. self-hosted models. It is specialized for decision tasks and cannot be used for general chat, text generation, or summarization. Performance is contingent on the quality and specificity of the user-provided "state" and question criteria.
2. Key Alternatives & Differentiation:
   • Jev 1.13 (Typesafe): A direct competitor in the decision model space. Nimble-v4-max claims a significant +10.5 point lead on the public Decision Index suite. Differentiation includes Nimble's hosted API service model versus Jev's more research-oriented, often self-hosted deployment.
   • Perplexity Decider v1.1: Another benchmark competitor. Nimble-v4-max leads by over 6 points. A key functional difference is Nimble's design for multiple questions per state, potentially offering a more efficient workflow for complex decision trees compared to single-question approaches.
   • Using General-Purpose LLMs (e.g., GPT-4, Claude): Developers often prompt large chat models for decisions. Nimble differentiates by being cheaper, faster, and purpose-built to return structured probabilities natively, eliminating prompt engineering for output formatting and reducing hallucination in structured tasks.

Frequently Asked Questions (FAQ)

1. What is the Nimble AI decision model? Nimble is a hosted API that provides specialized AI models which analyze text and return probabilistic scores for structured decisions—like classifying a support ticket or checking a policy—instead of generating conversational text.
2. How does Nimble compare to Jev? Based on the public Decision Index benchmark (v0.3.2), Nimble's models outperform Jev 1.13. Nimble-v4-max scores 68.46 (+10.5 vs. Jev) and Nimble-v4 scores 63.70 (+5.74 vs. Jev), making them top-ranked models, especially when considering their parameter size.
3. What are the main use cases for the Nimble API? Key use cases are AI workflow automation where clear decisions are needed: customer support routing and triage, content moderation, evaluation and grading of responses, and controlling the decision points within AI agent systems.
4. What is the difference between Nimble-v4 and Nimble-v4-max? Nimble-v4 is a 9-billion parameter model optimized for high-volume, low-latency tasks like routing and tagging. Nimble-v4-max is a 27-billion parameter model designed for higher-stakes decisions where maximum accuracy is critical, such as policy checks or grading.
5. How do I get started with the Nimble decision API? You can sign up for a free account at console.bespokelabs.ai, which includes credits for approximately 3,000 decisions. The API uses a straightforward JSON request format to send a "state" (context) and "questions" with typed criteria, returning JSON with probabilities for each answer.