glm4-moe-tiny

A small (~543M parameter) GLM-4 MoE model for testing prime-rl integration with HuggingFace Transformers and vLLM.

Purpose

This model is not for production use. It exists to:

• Validate MoE weight conversion between HuggingFace and PrimeRL formats
• Test the full RL training pipeline (inference server + trainer) at small scale
• Catch architecture-specific bugs without needing 100B+ parameter models

The model has been fine-tuned on PrimeIntellect/Reverse-Text-SFT to provide a non-trivial distribution for KL divergence during RL.

Quick Start

# Run RL with reverse-text environment
uv run rl @ configs/ci/integration/rl_moe/glm4_moe.toml

See the Testing MoE at Small Scale guide for full instructions.

Model Details

Parameter	Value
Hidden size	1024
Layers	24
Experts	8
Active experts	4
Parameters	~543M

Links

• prime-rl - RL training framework
• PrimeIntellect - Building infrastructure for decentralized AI