markopoloaiinc/Athena-mvrko-4B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 23, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

markopoloaiinc/Athena-mvrko-4B is a 4-billion-parameter decoder-only causal language model developed by Markopolo AI, fine-tuned from Qwen/Qwen3-4B. Optimized for predicting exact next actions of shoppers on retail pages, it achieves 24.50% strict exact-match on the OPeRA test set, outperforming larger frontier models like GPT-5.6 and Claude Sonnet 5. This model specializes in behavioral prediction by utilizing a learned, structure-preserving observation space and 32,768-token long-context supervision, making it ideal for agentic shopping and behavioral simulation.

Loading preview...

Overview

Athena (4B) is a specialized Large Event Model developed by Markopolo AI, fine-tuned from Qwen/Qwen3-4B. It is designed to predict the exact next action a real shopper will take on a live retail page, emitting structured JSON output. This 4-billion-parameter model achieves a 24.50% strict exact-match on the full official OPeRA test set (992 actions), ranking first among all models tested, including GPT-5.6, Claude Sonnet 5, and Claude Opus 4.8.

Key Capabilities

  • Superior Next-Action Prediction: Outperforms frontier general-purpose LLMs on the OPeRA benchmark for predicting shopper behavior, demonstrating the power of behavioral specialization over raw model scale.
  • Structured JSON Output: Emits next actions as schema-valid JSON, with 98.9% schema validity on the test set.
  • Long-Context Understanding: Trained with a 32,768-token context length, allowing it to condition on full interaction histories and large page states.
  • Efficient and Self-Hostable: As a 4B model, it is significantly more cost-effective for inference than larger API-based models, offering self-hostability and local inference capabilities.
  • Innovative Observation Space: Utilizes a learned, structure-preserving reduction of web page states, making complex DOM information model-legible and predictable.

Good For

  • Next-action prediction and autocomplete of shopper intent on e-commerce platforms.
  • Session replay and scoring logged user journeys step by step.
  • Serving as a foundation for agentic shopping and behavioral simulation in environments that can return new page states for novel actions.