renaudb1999/le-harnais-ft-agentworld-3b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 26, 2026License:llama3.2Architecture:Transformer Featherless Exclusive Cold

The renaudb1999/le-harnais-ft-agentworld-3b is a 3.2 billion parameter world-model student, built upon the Meta Llama-3.2-3B-Instruct architecture with a 32768 token context length. This model is specifically fine-tuned for agentic tasks, demonstrating a balanced quality-to-speed ratio. It achieves a token-F1 score of 0.868 and an OBSERVATION hit-rate of 75% against its teacher model, Qwen-AgentWorld-35B-A3B, making it suitable for applications requiring efficient world modeling.

Loading preview...

Model Overview

renaudb1999/le-harnais-ft-agentworld-3b is a 3.2 billion parameter language model designed as a "world-model student," prioritizing a balance between performance and inference speed. It is built on the Meta Llama-3.2-3B-Instruct base model, inheriting its architecture and subject to the Llama Community License.

Key Capabilities

  • Agentic World Modeling: Fine-tuned specifically for agentic tasks, distilling knowledge from a larger teacher model, Qwen-AgentWorld-35B-A3B.
  • Performance Metrics: Achieves a token-F1 score of 0.868 and an OBSERVATION hit-rate of 75% when compared to its teacher model on a 40-example evaluation set.
  • Efficient Inference: Optimized for scenarios where a balance between output quality and rapid processing is crucial.
  • Flexible Formats: Provided in multiple formats for diverse deployment needs:
    • .safetensors: For bf16 inference using transformers or le-harnais's lh-serve/candle.
    • .Q4_K_M.gguf: Portable 4-bit quantization, compatible with ollama / llama.cpp (Mac-friendly).
    • .Q8_0.gguf: Higher-fidelity 8-bit quantization for "hero models."

Good For

  • Agent-based Systems: Ideal for integrating into agentic workflows where a compact yet capable world model is required.
  • Resource-Constrained Environments: Suitable for deployment on hardware with limited resources, thanks to its 3.2B parameter count and optimized performance.
  • Rapid Prototyping: Its balanced speed and quality make it a strong candidate for quickly developing and testing agentic applications.

Finetunes
1 models