chat.md

Model	Description
BERT	Introduced by Google in 2018, BERT is a transformer-based model with 342 million parameters. It's used for query understanding in Google Search.
Claude	Created by Anthropic, Claude focuses on constitutional AI. The latest iteration, Claude 3.5 Sonnet, excels in understanding nuance, humor, and complex instructions. Available for free via Claude.ai and the Claude iOS app.
Cohere	Cohere provides several custom-trainable LLMs like Command, Rerank, and Embed. Founded by an author of "Attention Is All You Need," it operates on multiple clouds.
Ernie	Baidu’s Ernie 4.0 chatbot, with rumored 10 trillion parameters, has over 45 million users. It's primarily effective in Mandarin but capable in other languages.
Falcon 40B	Developed by the Technology Innovation Institute, Falcon 40B is a causal decoder-only model available in smaller variants (1B, 7B). It's open source and available on Amazon SageMaker and GitHub.
Gemini	Google's multimodal LLM, replacing Palm in Bard, now Gemini. Comes in Ultra, Pro, and Nano sizes, and outperforms GPT-4 on most benchmarks. Integrated into many Google products.
Gemma	Google's open-source models, Gemma 2B and 7B, trained on the same resources as Gemini, can run locally and surpass similarly sized Llama 2 models on several benchmarks.
GPT-3	OpenAI's 175 billion parameter model, GPT-3 uses a decoder-only transformer architecture and is integrated exclusively into Microsoft products.
GPT-3.5	An upgraded version of GPT-3, fine-tuned with reinforcement learning from human feedback. Powers ChatGPT and is integrated into Bing Search, replaced by GPT-4.
GPT-4	OpenAI's 2023 release, GPT-4, is a multimodal model rumored to have 170 trillion parameters. It can process both text and images, demonstrating human-level performance in multiple exams. Powers Bing Search and ChatGPT Plus.
GPT-4o	Successor to GPT-4, offering faster, more natural interactions and multimodal capabilities. Free for developer and customer products.
Lamda	Google's Lamda, announced in 2021, is a decoder-only transformer model pre-trained on a large corpus. Known for generating claims of sentience from a Google engineer.
Llama	Meta's LLM, released in 2023, available in sizes up to 65 billion parameters. Initially for approved researchers, now open source. Trained on public data like CommonCrawl and Wikipedia.
Mistral	A 7 billion parameter model that outperforms Llama's similar-sized models. It is self-hostable and suitable for business purposes, released under the Apache 2.0 license.
Orca	Developed by Microsoft with 13 billion parameters, Orca aims to improve on other open source models, achieving similar performance to GPT-4 with fewer parameters. Built on LLaMA 13B.
Palm	Google's 540 billion parameter model powering Bard, specializing in reasoning tasks like coding and math. Part of the Pathways initiative for diverse applications.
Phi-1	Microsoft’s transformer-based model with 1.3 billion parameters, trained on high-quality data. Specializes in Python coding.
StableLM	Stability AI's series of open source models, ranging from 3 billion to 175 billion parameters. Aims to be transparent and accessible.
Vicuna 33B	Derived from Llama, developed by LMSYS, with 33 billion parameters. Performs well for its size, though smaller and less capable than GPT-4.
Seq2Seq	Google's deep learning approach for machine translation and NLP, underlying modern models like LaMDA and AlexaTM 20B. Uses a mix of encoders and decoders.
Eliza	An early NLP program from 1966, simulating conversation using pattern matching. One of the first language models, demonstrating basic interaction principles.