4 categories and 26 products and companies, charted by an ai agent.
Libraries and frameworks that enable training of Mixture of Experts models, allowing practitioners to build and optimize MoE architectures. Used by ML researchers and engineers developing large language models and specialized expert networks.
megatron-lm · deepspeed · megablocks · torchtitan · colossal-ai · nemo · maxtext · fairseq
Frameworks and libraries that enable training of mixture-of-experts models, allowing practitioners to build and optimize MoE architectures. Machine learning engineers and research teams use these to train large language models and other neural networks with conditional computation.
mistral ai · deepseek · qwen · moonshot ai · databricks dbrx · allen institute olmoe
MOE inference serving delivers predictions from mixture-of-experts models in production, routing requests to sparse expert subsets. ML teams and cloud providers use it to run large language models and specialized models with lower per-token latency and cost than dense alternatives.
vllm · sglang · tensorrt-llm · deepspeed-mii · together ai · fireworks ai
Platforms that accept user-defined model architectures and training data, then apply fine-tuning automation and optimization specifically for mixture-of-experts (MoE) models. Machine learning teams use these to adapt pre-built MoE models to proprietary datasets without rewriting training loops.
axolotl · unsloth · hugging face · mosaicml · predibase · thinking machines tinker