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

Dense decoder transformers

The GPT design: one stack of transformer layers predicting the next token. 308 models on EveryLLM so far.

The dense decoder transformer is the design behind GPT and most chatbots. It is a single stack of identical transformer layers that reads the tokens so far and predicts the next one, with a mask that stops each position from looking ahead. Dense means every parameter is used for every token, unlike in a mixture of experts. GPT-3, PaLM, Llama 3, Chinchilla, Gemma, Phi and OLMo are dense decoders.

Inside each layer

Each layer has two parts. Self-attention lets every token weigh every earlier token and gather what it needs; a feed-forward network then transforms each token's vector on its own. The transformer was introduced by Google researchers in the June 2017 paper Attention Is All You Need, for translation, and OpenAI's GPT-1 (2018) kept only its decoder half. Llama 3.1 405B, released in July 2024, stacks 126 such layers.

Why it won

A decoder trains on any text without labels, because the next token is always the answer to predict, and it scales smoothly: scaling laws published by OpenAI in 2020 and DeepMind in 2022 showed the error falling predictably with more parameters, data and compute. Its weak point is cost. Attention grows with the square of the context length, and every parameter must be read for every token, which is why mixtures of experts became popular.

Famous dense decoder models

Still to be written: OLMo (language model).

Further reading

fnInternLM2 馃嚚馃嚦 ChinaInternLM2 Open-source 20 billion parameter language model with long context supportGeneral 路 Open weights 路 2024 Meltemi 馃嚞馃嚪 GreeceMeltemi Open large language model for GreekGeneral 路 Open weights 路 2024 wavVoxtral 馃嚝馃嚪 FranceVoxtral Open audio chat models that transcribe, answer questions and summarise speechMultimodal 路 Open weights 路 2025 Mi:dm 2.0 馃嚢馃嚪 South KoreaMi:dm 2.0 Korea-centric language model by KTGeneral 路 Open weights 路 2025 DeepSeek-V2 馃嚚馃嚦 ChinaDeepSeek-V2 Efficient mixture-of-experts language model with 236 billion parametersGeneral 路 Open weights 路 2024 CyberAgentLM3 馃嚡馃嚨 JapanCyberAgentLM3 A 22-billion parameter Japanese and English chat model by CyberAgent.General 路 Open weights 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 Falcon-7B 馃嚘馃嚜 United Arab EmiratesFalcon-7B Open 7-billion parameter decoder model by TIIGeneral 路 Open weights 路 2023 fnGranite 4.2 馃嚭馃嚫 United StatesGranite 4.2 IBM's open reasoning language model family, released August 2026Reasoning 路 Open weights 路 2026 ALLaM 馃嚫馃嚘 Saudi ArabiaALLaM Bilingual Arabic and English language model seriesGeneral 路 Open weights 路 2024 Phi-2 馃嚭馃嚫 United StatesPhi-2 A 2.7 billion parameter Microsoft research model for QA, chat and codeResearch 路 Open weights Gemma 2 馃嚭馃嚫 United StatesGemma 2 Open lightweight language models from Google DeepMind, 2B to 27B parametersGeneral 路 Open weights 路 2024 imgvidfnGemma 3 馃嚭馃嚫 United StatesGemma 3 Lightweight multimodal open models from 1B to 27B parametersGeneral 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 Llama 2 馃嚭馃嚫 United StatesLlama 2 Meta's open family of pretrained and chat-tuned language models, 7B to 70BGeneral 路 Open weights 路 2023 Phi-4 馃嚭馃嚫 United StatesPhi-4 A 14B Microsoft Research model trained heavily on synthetic data for reasoning.General 路 Open weights 路 2024 fnBlenderBot 3 WorldBlenderBot 3 A 175B open-domain chatbot with internet access and long-term memoryGeneral 路 Open weights 路 2022 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 Phi-4-reasoning 馃嚭馃嚫 United StatesPhi-4-reasoning A 14B open-weight reasoning model fine-tuned from Phi-4Reasoning 路 Open weights 路 2025 Falcon2-11B 馃嚘馃嚜 United Arab EmiratesFalcon2-11B An 11 billion parameter causal decoder model by TII.General 路 Open weights 路 2024 DeepSeek-R1 馃嚚馃嚦 ChinaDeepSeek-R1 Open-weight reasoning model from ChinaReasoning 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.