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

wavfnSarvam-30B 馃嚠馃嚦 IndiaSarvam-30B Open-weight MoE model for Indian languages and codingGeneral 路 Open weights PanGu-Sigma 馃嚚馃嚦 ChinaPanGu-Sigma Trillion-parameter sparse language model for Chinese NLP tasksGeneral 路 Research only 路 2023 fnGLM-5 馃嚚馃嚦 ChinaGLM-5 Open-weight model for engineering and agentic tasksGeneral 路 Open weights 路 2026 imgERNIE 4.5 馃嚚馃嚦 ChinaERNIE 4.5 Baidu's multimodal mixture-of-experts language model series.General 路 Open weights 路 2025 fnHunyuan-A13B 馃嚚馃嚦 ChinaHunyuan-A13B Open-source MoE language model with 80 billion total parametersGeneral 路 Open weights 路 2025 Hunyuan-TurboS 馃嚚馃嚦 ChinaHunyuan-TurboS Hybrid Mamba-Transformer MoE model by TencentGeneral 路 API only 路 2025 Seed1.5-Thinking 馃嚚馃嚦 ChinaSeed1.5-Thinking A reasoning model from ByteDance Seed using reinforcement learningReasoning 路 API only 路 2025 MiMo-V2.5-Pro 馃嚚馃嚦 ChinaMiMo-V2.5-Pro Open-source MoE language model with 1M token contextGeneral 路 Open weights imgwavvidMiMo-V2.6-Pro 馃嚚馃嚦 ChinaMiMo-V2.6-Pro Xiaomi鈥檚 1T parameter multimodal model with 1M token contextMultimodal 路 Open weights fnStep-3.5-Flash 馃嚚馃嚦 ChinaStep-3.5-Flash Open-source MoE model for reasoning and agentsGeneral 路 Open weights 路 2026 MAI-1-preview 馃嚭馃嚫 United StatesMAI-1-preview Microsoft AI's first in-house foundation model, a mixture-of-experts system for everyday queriesGeneral 路 API only 路 2025 Qwen2 馃嚚馃嚦 ChinaQwen2 Open-weight language model series from AlibabaGeneral 路 Open weights 路 2024 fnK-EXAONE 馃嚢馃嚪 South KoreaK-EXAONE LG AI Research multilingual Mixture-of-Experts modelGeneral 路 Open weights fnGranite 4.0 馃嚭馃嚫 United StatesGranite 4.0 IBM's hybrid Mamba/transformer language models for enterprise workGeneral 路 Open weights 路 2025 imgfnMistral Large 3 馃嚝馃嚪 FranceMistral Large 3 A 675B-parameter open-weight mixture-of-experts multimodal model from Mistral AIGeneral 路 Open weights imgApple Foundation Models 馃嚭馃嚫 United StatesApple Foundation Models Apple's 2025 on-device and server language models for Apple IntelligenceGeneral 路 API only 路 2025 fnSolar Open 2 馃嚢馃嚪 South KoreaSolar Open 2 250B parameter open-weight model for agentic tasksGeneral 路 Open weights 路 2026 Mixtral 8x22B 馃嚝馃嚪 FranceMixtral 8x22B Sparse mixture-of-experts open model with 39B active parameters out of 141BGeneral 路 Open weights fnGpt-oss-120b 馃嚭馃嚫 United StatesGpt-oss-120b OpenAI's open-weight reasoning model that fits on one 80GB GPUReasoning 路 Open weights 路 2025 imgwavvidGemini 1.5 Pro 馃嚭馃嚫 United StatesGemini 1.5 Pro Google DeepMind's mid-size multimodal model with up to a 1 million token context windowGeneral 路 API only 路 2024 DeepSeek-V2 馃嚚馃嚦 ChinaDeepSeek-V2 Efficient mixture-of-experts language model with 236 billion parametersGeneral 路 Open weights 路 2024 fnGpt-oss-20b 馃嚭馃嚫 United StatesGpt-oss-20b OpenAI's smaller open-weight reasoning model for local and low-latency useReasoning 路 Open weights 路 2025 imgwavvidGemma 4 馃嚭馃嚫 United StatesGemma 4 Google DeepMind's open multimodal models in five sizes, from phones to serversGeneral 路 Open weights GLaM 馃嚭馃嚫 United StatesGLaM A trillion-parameter mixture-of-experts language model from Google ResearchResearch 路 Research only 路 2021 fnNemotron 3 Super 馃嚭馃嚫 United StatesNemotron 3 Super NVIDIA's 120B hybrid MoE language model with 12B active parameters and 1M-token contextGeneral 路 Conditional licence 路 2026 imgvidfnStep-5-Preview WorldStep-5-Preview Agentic model with 600B parameters and 1M token contextGeneral 路 Open weights 路 2026 Granite 3.0 馃嚭馃嚫 United StatesGranite 3.0 IBM's third-generation open enterprise language models, from 2B to 8B parametersGeneral 路 Open weights 路 2024 Grok-1 馃嚭馃嚫 United StatesGrok-1 xAI's 314 billion parameter open-weights mixture-of-experts base modelGeneral 路 Open weights 路 2024 Sarvam-105B 馃嚠馃嚦 IndiaSarvam-105B Open-weight MoE model for Indian languages and reasoningGeneral 路 Open weights Hunyuan-Large 馃嚚馃嚦 ChinaHunyuan-Large Open-source mixture of experts model with 389 billion parametersGeneral 路 Open weights 路 2024 Param-1 馃嚠馃嚦 IndiaParam-1 Bilingual and multilingual language models from India.General 路 Conditional licence fnMiniMax-M2 馃嚚馃嚦 ChinaMiniMax-M2 Open-source MoE model for coding and agentsCode 路 Open weights fnNemotron 3 Nano 馃嚭馃嚫 United StatesNemotron 3 Nano NVIDIA's 30B hybrid Mamba-Transformer MoE model for reasoning and agentsGeneral 路 Conditional licence 路 2025 imgvidfnKimi K3 馃嚚馃嚦 ChinaKimi K3 Open-weight multimodal agentic model with 2.8T parametersGeneral 路 Open weights 路 2026 OLMoE 馃嚭馃嚫 United StatesOLMoE Open mixture-of-experts language model with 1 billion active parametersGeneral 路 Fully open 路 2024 DeepSeek-R1 馃嚚馃嚦 ChinaDeepSeek-R1 Open-weight reasoning model from ChinaReasoning 路 Open weights 路 2024 Jamba 馃嚠馃嚤 IsraelJamba Hybrid open-weights language model with long contextGeneral 路 Open weights 路 2024

The weekly digest

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