News
ReleasesOVHcloud 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 improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud 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 improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
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

1.58-bit large language model World1.58-bit large language model A model whose weights hold only three values: -1, 0 and +1.Research · Open weights · 2024 Chinchilla 🇬🇧 United KingdomChinchilla A 70-billion-parameter DeepMind model built to test how models should scale.Research · Research only · 2022 Stanford Alpaca 🇺🇸 United StatesStanford Alpaca A cheap instruction-tuned LLaMA 7B built at Stanford in 2023Research · Research only · 2023 imgwavvidVideoPoet 🇺🇸 United StatesVideoPoet A Google Research language model that turns text and images into video and audioMultimodal · Research only · 2023 fnSparrow 🇬🇧 United KingdomSparrow DeepMind's rule-following English chatbot that cites web evidence for its answersAssistant · Research only · 2022 Gopher 🇬🇧 United KingdomGopher A 280 billion parameter Transformer language model from DeepMind, studied for how scale changes ability.Research · Research only · 2021 DialoGPT 🇺🇸 United StatesDialoGPT A GPT-style model trained on Reddit threads to write chat repliesGeneral · Open weights · 2019 OPT 🇺🇸 United StatesOPT Meta AI's family of open decoder-only language models from 125M to 175B parametersGeneral · Conditional licence · 2022 AlphaCode 🇬🇧 United KingdomAlphaCode A code generation system that solves competitive programming problems.Code · Research only · 2022 RETRO WorldRETRO A retrieval-enhanced transformer that conditions on chunks from a huge document database.Research · Research only · 2021 imgMM1 WorldMM1 A family of multimodal language models up to 30B parameters, dense and MoE.Multimodal · Research only · 2024 fnToolformer 🇺🇸 United StatesToolformer A language model that learns on its own when to call external tools through APIs.Research · Research only · 2023 Orca 2 🇺🇸 United StatesOrca 2 A small LLaMA-2-based model trained on synthetic data to reason.Small · Conditional licence · 2023 fnWebGPT 🇺🇸 United StatesWebGPT A GPT-3 model fine-tuned to answer questions while browsing the webGeneral · Research only · 2021 Med-PaLM 🇺🇸 United StatesMed-PaLM A medical-domain language model aligned from Flan-PaLM using instruction prompt tuning.Domain · Research only · 2022 Minerva 🇺🇸 United StatesMinerva A language model trained further on technical content to solve maths and science problemsResearch · Research only · 2022

The weekly digest

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