stay-mellow-ai/mev

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The mev model by Mellow AI is a 4.5 billion parameter decision model, fine-tuned from Qwen3.5-4B. It is specifically designed to choose one option from a structured list given a state and a question, making it highly effective for automated decision-making tasks. This model excels at structured classification and rule application, providing a single letter response for its chosen option. It offers reliable confidence metrics for its decisions, making it suitable for applications requiring high accuracy in automated choices.

Loading preview...

Mellow AI's mev: A Specialized Decision Model

mev is a 4.5 billion parameter model developed by Mellow AI, built upon a supervised LoRA fine-tune of Qwen3.5-4B. Unlike general-purpose language models, mev is engineered for structured decision-making, taking a state, a question, and a list of options to return a single, definitive choice.

Key Capabilities

  • Structured Decision-Making: Processes inputs with a state (text or JSON), a question, and 2 to 24 labeled options.
  • Single Option Output: Returns precisely one option letter, designed for clear, unambiguous choices.
  • Confidence Metrics: Provides client-side computed confidence scores for its decisions, with an expected calibration error of 0.038 and an AUROC of 0.88 on its development set. This allows for setting thresholds for automated action versus human review.
  • Typed Question Support: Can be adapted to answer choice, noul (yes/no), and score type questions by mapping them to its decision format.
  • Optimized for Specific Tasks: Trained on a diverse dataset including MultiNLI, BoolQ, Banking77, AG News, SST-5, and synthetic rule application tasks.

When to Use mev

  • Automated Classification: Ideal for tasks requiring the model to classify inputs into predefined categories, such as sentiment analysis, intent recognition, or news topic classification.
  • Rule Application: Excellent for scenarios where a decision needs to be made based on a set of rules, like determining if a return is within policy.
  • High-Confidence Decisions: When applications require a high degree of certainty in automated choices, mev's confidence metrics allow for strategic implementation, such as routing low-confidence decisions to human review.
  • Resource-Efficient Decision Logic: As a 4.5B parameter model, it offers a more efficient solution for specific decision tasks compared to larger, more general-purpose LLMs, especially when deployed locally via vLLM.

Finetunes
1 models