Jainamshahhh/parry-tactician-1.5b-merged

TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 10, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

Jainamshahhh/parry-tactician-1.5b-merged is a 1.5 billion parameter language model designed as the opponent brain for the Parry real-time 1v1 footsies duel game. This model generates grammar-constrained intent letters, conditioned on a natural-language plan-string, and operates efficiently in-browser via GGUF Q4_K_M conversion for WASM runtimes. It is specifically optimized for plan-conditioned action generation in a game environment, demonstrating robust steering capabilities through classifier-free guidance.

Loading preview...

Model Overview

Jainamshahhh/parry-tactician-1.5b-merged is a 1.5 billion parameter model developed as the "opponent brain" for the Parry real-time 1v1 footsies duel game. Its core function is to generate single-letter, grammar-constrained intent actions (e.g., S F P L R W) approximately every 160ms. A key differentiator is its ability to be conditioned on a natural-language plan-string, which directly steers its behavior. This plan can be edited live, allowing for dynamic control over the model's actions.

Key Capabilities & Training

  • Plan-Conditioned Action Generation: The model's output is directly influenced by a natural-language plan, enabling precise control over its in-game strategy.
  • In-Browser Execution: Converted to GGUF Q4_K_M, it runs efficiently in-browser using llama.cpp's WASM runtime (wllama), acting as an "Analyst" to generate and interpret plans.
  • Classifier-Free Guidance (CFG): Training included 20% plan-dropout, allowing for CFG at inference. This technique significantly enhances the model's obedience to the conditioning plan, achieving a ΔP of 0.292 at γ=4, which is 88% of the expert ceiling.
  • Behavior Cloning: Trained using behavior cloning (LoRA r=32 α=64) on plan-conditioned duel logs, incorporating on-trajectory expert play and counterfactual state-plan expansions.

When to Use This Model

This model is specifically designed for applications requiring real-time, plan-conditioned action generation within constrained environments, particularly for game AI or interactive agents where explicit strategic guidance is beneficial. Its in-browser capability makes it suitable for web-based applications where local execution is preferred.