🧨 FLAME-MoE
FLAME-MoE is a fully open Mixture-of-Experts (MoE) language model suite developed by Carnegie Mellon University. It provides a transparent and reproducible research platform for investigating expert routing, model scaling, and training dynamics in sparse architectures. The suite includes seven decoder-only transformer models ranging from 38M to 1.7B active parameters and reflects production-grade MoE setups with 64 experts per MoE layer, top-8 routing, and shared experts.
🔍 Model Summary
| Model Name | Active / Total Params | Layers | MoE Experts (Total/Active/Shared) | Training FLOPs | Tokens Trained |
|---|
| FLAME-MoE-38M-100M | 38M / 100M | 9 | 64 / 8 / 2 | 1.0e18 | 4.4B |
| FLAME-MoE-98M-349M | 98M / 349M | 9 | 64 / 8 / 2 | 3.0e18 | 5.0B |
| FLAME-MoE-115M-459M | 115M / 459M | 12 | 64 / 8 / 2 | 6.0e18 | 8.7B |
| FLAME-MoE-290M-1.3B | 290M / 1.3B | 9 | 64 / 8 / 2 | 2.0e19 | 11.4B |
| FLAME-MoE-419M-2.2B | 419M / 2.2B | 15 | 64 / 8 / 2 | 3.0e19 | 11.9B |
| FLAME-MoE-721M-3.8B | 721M / 3.8B | 12 | 64 / 8 / 2 | 8.0e19 | 18.4B |
| FLAME-MoE-1.7B-10.3B | 1.7B / 10.3B | 18 | 64 / 8 / 2 | 2.4e20 | 23.1B |
📖 Training Details
- Framework: Megatron-LM with Expert Parallelism (EP=8), Pipeline Parallelism (PP=1)
- Data: Pretrained on DataComp-LM (DCLM)
- Batch Size: 1024
- Sequence Length: 2048
- Optimizer: Adam
- Scheduler: WSD (Warmup + Decay)
- Learning Rate: Max 3e-4, Min 3e-5
- Checkpoints: 10 saved per model across training
- Hardware: 32× NVIDIA H100 GPUs
🛠 Intended Use
FLAME-MoE is developed for research purposes only. It supports academic study in:
- Sparse model training dynamics
- Expert routing behavior and specialization
- Scaling laws and compute-optimal design
- Benchmarking and reproducibility in MoE LLMs
It is not intended for commercial deployment or instruction-tuned downstream tasks.
📂 Access
All models, training scripts, logs, routing traces, and evaluation pipelines are available at:
