News
EveryLLM

Mixture of experts

Mixtral, DeepSeek-V3, Kimi K2: many expert networks, a few used for each token. 112 models on EveryLLM so far.

A mixture of experts (MoE) model replaces each dense feed-forward block with many smaller expert networks and a router that sends each token to only a few of them. The model can hold a very large number of parameters while using only a fraction for any one token, so it learns more without costing proportionally more to run. Mixtral, DeepSeek-V3, Qwen3, Llama 4, gpt-oss and Kimi K2 are all built this way.

Total and active parameters

An MoE has two sizes. DeepSeek-V3 (December 2024) holds 671 billion parameters but activates 37 billion for each token. Mixtral 8x7B (December 2023) has eight experts per layer, uses two for each token, and has 46.7 billion parameters in total with 12.9 billion active. Kimi K2 (July 2025) reaches one trillion in total with 32 billion active. All the experts must still sit in memory, so MoE models need large servers even when they are fast.

A long history

The idea dates to 1991, and Google researchers led by Noam Shazeer revived it for language in 2017 with a sparsely gated layer of thousands of experts. Google's Switch Transformer reached 1.6 trillion parameters in 2021. Open MoE models took off with Mixtral, then Databricks' DBRX (March 2024, 132 billion total, 36 billion active), and became the norm for large Chinese open models. Training them is harder, because routers must spread tokens evenly across experts.

Famous mixture of experts models

Further reading

imgfnMistral Large 3 🇫🇷 FranceMistral Large 3 A 675B-parameter open-weight mixture-of-experts multimodal model from Mistral AIGeneral · Open weights fnSolar Open 2 🇰🇷 South KoreaSolar Open 2 250B parameter open-weight model for agentic tasksGeneral · Open weights · 2026 Mixtral 8x22B 🇫🇷 FranceMixtral 8x22B Sparse mixture-of-experts open model with 39B active parameters out of 141BGeneral · Open weights imgvidfnStep-5-Preview WorldStep-5-Preview Agentic model with 600B parameters and 1M token contextGeneral · Open weights · 2026 fnK-EXAONE 🇰🇷 South KoreaK-EXAONE LG AI Research multilingual Mixture-of-Experts modelGeneral · Open weights fnStep-3.5-Flash 🇨🇳 ChinaStep-3.5-Flash Open-source MoE model for reasoning and agentsGeneral · Open weights · 2026 Qwen2 🇨🇳 ChinaQwen2 Open-weight language model series from AlibabaGeneral · Open weights · 2024 Grok-1 🇺🇸 United StatesGrok-1 xAI's 314 billion parameter open-weights mixture-of-experts base modelGeneral · Open weights · 2024 Sarvam-105B 🇮🇳 IndiaSarvam-105B Open-weight MoE model for Indian languages and reasoningGeneral · Open weights Hunyuan-Large 🇨🇳 ChinaHunyuan-Large Open-source mixture of experts model with 389 billion parametersGeneral · Open weights · 2024 fnMiniMax-M2 🇨🇳 ChinaMiniMax-M2 Open-source MoE model for coding and agentsCode · Open weights imgvidfnKimi K3 🇨🇳 ChinaKimi K3 Open-weight multimodal agentic model with 2.8T parametersGeneral · Open weights · 2026 DeepSeek-R1 🇨🇳 ChinaDeepSeek-R1 Open-weight reasoning model from ChinaReasoning · Open weights · 2024 Jamba 🇮🇱 IsraelJamba Hybrid open-weights language model with long contextGeneral · Open weights · 2024

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.