Models Qwen 30b6

0xgr3y/Qwen3-0.6B-Gensyn-Swarm-tall_tame_panther

Cold
Public
0.8B
BF16
40960
License: apache-2.0
Hugging Face
Overview

Model Overview

This model, named tall_tame_panther, is a continuously trained Qwen3-0.6B variant developed by 0xgr3y. It is fine-tuned using the Gensyn RL-Swarm framework with GRPO (Generalized Reward Policy Optimization), focusing specifically on enhancing reasoning and mathematical capabilities. The model is in LIVE training, with automatic updates every 5-10 minutes, and supports GGUF (llama.cpp) for broad compatibility.

Key Features

  • Real-time Training: Undergoes continuous learning through a distributed RL process across the Gensyn swarm network.
  • Multi-domain Reasoning: Specialized in logic, mathematical problem-solving, and various reasoning tasks.
  • GGUF Support: Offers multiple quantized formats (F16, Q3_K_M, Q4_K_M, Q5_K_M) for efficient deployment.
  • llama.cpp & TGI Compatible: Ready for edge deployment, local inference, and production use with Text Generation Inference.
  • BF16 Precision: Trained using bfloat16 for optimal performance.
  • Qwen3 Chat Format: Inherits the Qwen3 chat template for conversational interactions, though current training prioritizes math/reasoning.

Training Details

The model is trained on a composite dataset of 1,000 samples, with a weighted sampling strategy emphasizing propositional logic, calendar arithmetic, decimal arithmetic, and base conversion. Training is ongoing, with over 43,610 rounds completed out of 100,000, and weights are updated frequently. The decentralized RL architecture provides fault tolerance and collective intelligence.

Good For

  • Logical Reasoning: Tasks involving propositional logic, truth evaluation, and Boolean algebra.
  • Mathematical Operations: Multi-precision arithmetic, decimal calculations, fractions, and number system conversions.
  • Step-by-step Problem Solving: Excels at chain-of-thought reasoning for complex problems.
  • Conversational Tutoring: Can be used for interactive problem-solving in a chat format.
  • Edge Deployment: Its smaller size (0.6B parameters) and GGUF support make it suitable for local inference on resource-constrained devices.