renaudb1999/le-harnais-ft-agentworld-8b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 26, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

The renaudb1999/le-harnais-ft-agentworld-8b is an 8 billion parameter language model built upon Meta-Llama-3.1-8B-Instruct, designed as a world-model student. It achieves near teacher-quality performance, specifically a token-F1 of 0.958 and an OBSERVATION hit-rate of 95% against its Qwen-AgentWorld-35B-A3B teacher. This model is optimized for agentic tasks, providing a capable local replacement for larger teacher models at approximately 16GB. Its primary strength lies in its ability to mimic complex agent behaviors efficiently.

Loading preview...

renaudb1999/le-harnais-ft-agentworld-8b: Agentic World-Model Student

This model, le-harnais-ft-agentworld-8b, is an 8 billion parameter language model derived from meta-llama/Meta-Llama-3.1-8B-Instruct. It functions as a "world-model student," aiming to replicate the performance of a larger teacher model, specifically Qwen-AgentWorld-35B-A3B, for agentic tasks.

Key Capabilities & Performance

  • Near Teacher-Quality: Achieves a token-F1 score of 0.958 and an OBSERVATION hit-rate of 95% when evaluated against its teacher model on a dataset of 40 examples.
  • Efficient Agentic Behavior: Designed to provide a local, efficient replacement for more powerful teacher models, operating effectively within approximately 16GB of memory.
  • Llama-3.1 Foundation: Benefits from the robust architecture and capabilities of the Llama 3.1 instruction-tuned base model.

Formats Available

  • .safetensors: For bf16 inference, compatible with transformers or le-harnais serving tools.
  • .Q4_K_M.gguf: A portable 4-bit quantization suitable for ollama or llama.cpp, including Mac environments.
  • .Q8_0.gguf: A higher-fidelity 8-bit quantization for enhanced performance in hero model applications.

Use Cases

This model is particularly well-suited for applications requiring efficient, agent-like reasoning and interaction, where a smaller, locally deployable model can approximate the performance of larger, more resource-intensive alternatives. Its strong performance in token-F1 and observation hit-rate suggests its utility in scenarios demanding accurate and context-aware responses in agentic workflows.

Finetunes
1 models