News
ResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI鈥檚 New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI鈥檚 New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAI
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

imgwavvidGemma 馃嚭馃嚫 United StatesGemma Google DeepMind's open family of Gemini-derived language modelsGeneral 路 Open weights 路 2024 imgwavInkling WorldInkling An open-weights mixture-of-experts model handling text, images and audio inputGeneral 路 Open weights 路 2026 fnGLM 馃嚚馃嚦 ChinaGLM Open-weight Chinese model series used for chat and codingGeneral 路 Open weights 路 2021 imgwavvidQwen 馃嚚馃嚦 ChinaQwen A family of large and small language models developed by Alibaba Cloud.General 路 Open weights 路 2023 imgHuawei PanGu 馃嚚馃嚦 ChinaHuawei PanGu Huawei's multimodal large model family for industry, from 1B to trillion-scaleMultimodal 路 Open weights 路 2023 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 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 imgPhi-3 馃嚭馃嚫 United StatesPhi-3 Small Microsoft language models built to run well on phones and laptops.Small 路 Open weights 路 2024 DeepSeek-V3 馃嚚馃嚦 ChinaDeepSeek-V3 Open-weight mixture of experts language model with 671 billion parameters.General 路 Open weights 路 2024 fnKimi K2 Thinking 馃嚚馃嚦 ChinaKimi K2 Thinking Open-source reasoning model with tool use and INT4 quantizationReasoning 路 Open weights ERNIE 5.5 Flash 馃嚚馃嚦 ChinaERNIE 5.5 Flash Efficient sparse MoE language model with 280B total parametersGeneral 路 Open weights MiMo-V2-Flash 馃嚚馃嚦 ChinaMiMo-V2-Flash High-speed MoE language model for reasoning and agentsGeneral 路 Open weights PLLuM 馃嚨馃嚤 PolandPLLuM Polish-focused large language model familyGeneral 路 Open weights 路 2024 Snowflake Arctic 馃嚭馃嚫 United StatesSnowflake Arctic Open enterprise language model with hybrid MoE designGeneral 路 Open weights 路 2024 Mixtral 8x7B 馃嚝馃嚪 FranceMixtral 8x7B Sparse mixture-of-experts language model with open weightsGeneral 路 Open weights 路 2023 fnParam-2-17B 馃嚠馃嚦 IndiaParam-2-17B Multilingual MoE model for Indian languages and reasoningGeneral 路 Open weights DeepSeekMoE 馃嚚馃嚦 ChinaDeepSeekMoE Mixture-of-experts language model architecture by DeepSeek AIGeneral 路 Open weights 路 2024 DeepSeek-Coder-V2 馃嚚馃嚦 ChinaDeepSeek-Coder-V2 Open-source mixture-of-experts model for code and mathCode 路 Open weights 路 2024 fnTrinity Large 馃嚭馃嚫 United StatesTrinity Large 398B-parameter sparse MoE model family by Arcee AIGeneral 路 Open weights 路 2026 fnDeepSeek-V3.2 馃嚚馃嚦 ChinaDeepSeek-V3.2 Efficient reasoning and agentic AI model with sparse attentionReasoning 路 Open weights 路 2025 fnDeepSeek-V3.1 馃嚚馃嚦 ChinaDeepSeek-V3.1 Hybrid thinking and non-thinking language model with 671 billion parameters.General 路 Open weights fnDeepSeek-V4-Pro 馃嚚馃嚦 ChinaDeepSeek-V4-Pro 1.6T parameter MoE model with one million token contextGeneral 路 Open weights 路 2026 DeepSeek-V4-Flash 馃嚚馃嚦 ChinaDeepSeek-V4-Flash Efficient mixture-of-experts model with one million token contextGeneral 路 Open weights 路 2026 Qwen1.5 馃嚚馃嚦 ChinaQwen1.5 Open-weight language model series with nine sizes and strong multilingual support.General 路 Open weights 路 2024 fnK-EXAONE 2.0 馃嚢馃嚪 South KoreaK-EXAONE 2.0 Open-weight multilingual language model by LG AI ResearchGeneral 路 Open weights fnQwen3-235B-A22B 馃嚚馃嚦 ChinaQwen3-235B-A22B Mixture-of-experts model with switchable reasoning modesGeneral 路 Open weights 路 2025 A.X K1 馃嚢馃嚪 South KoreaA.X K1 Sparse MoE model with controllable reasoning depthGeneral 路 Open weights 路 2026 fnQwen3-Next-80B-A3B 馃嚚馃嚦 ChinaQwen3-Next-80B-A3B An 80-billion parameter hybrid model with ultra-long context support.General 路 Open weights imgfnQwen3.5-397B-A17B 馃嚚馃嚦 ChinaQwen3.5-397B-A17B A large multimodal language model with sparse activation.General 路 Open weights fnKanana-2 馃嚢馃嚪 South KoreaKanana-2 Open-source multilingual model for agentic AI and reasoning.General 路 Open weights fnKimi K2 馃嚚馃嚦 ChinaKimi K2 Open mixture-of-experts model for agentic tasks and codingGeneral 路 Open weights 路 2025 imgvidfnKimi K2.6 馃嚚馃嚦 ChinaKimi K2.6 Open-source multimodal agentic model by Moonshot AIGeneral 路 Open weights fnGLM-4.5 馃嚚馃嚦 ChinaGLM-4.5 Open-source mixture of experts model for agents and reasoningGeneral 路 Open weights 路 2025 Kimi Linear 馃嚚馃嚦 ChinaKimi Linear Hybrid linear attention model with 48 billion parametersGeneral 路 Open weights 路 2025 imgfnLing-1T 馃嚚馃嚦 ChinaLing-1T Trillion-parameter efficient reasoning model by InclusionAIGeneral 路 Open weights fnLongCat-2.0 馃嚚馃嚦 ChinaLongCat-2.0 A 1.6 trillion parameter MoE language model from Meituan.General 路 Open weights MiniMax-Text-01 馃嚚馃嚦 ChinaMiniMax-Text-01 A large language model with a four million token context window.General 路 Open weights 路 2025 wavfnSarvam-30B 馃嚠馃嚦 IndiaSarvam-30B Open-weight MoE model for Indian languages and codingGeneral 路 Open weights fnLongCat-Flash 馃嚚馃嚦 ChinaLongCat-Flash Efficient 560B parameter MoE model for agentic tasksGeneral 路 Open weights 路 2025 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 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 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

The weekly digest

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