Models
Cohere has a variety of models that cover many different use cases. If you need more customization, you can train a model to tune it to your specific use case.
Command
Command is Cohere's default generation model that takes a user instruction (or command) and generates text following the instruction. Our Command models also have conversational capabilities which means that they are well-suited for chat applications.
| Latest Model | Description | Max Tokens (Context Length) | Endpoints |
|---|---|---|---|
command-light | A smaller, faster version of command. Almost as capable, but a lot faster. | 4096 | Chat, Summarize |
command-light-nightly | To reduce the time between major releases, we put out nightly versions of command models. For command-light, that is command-light-nightly.Be advised that command-light-nightly is the latest, most experimental, and (possibly) unstable version of its default counterpart. Nightly releases are updated regularly, without warning, and are not recommended for production use. | 8192 | Chat |
command | An instruction-following conversational model that performs language tasks with high quality, more reliably and with a longer context than our base generative models. | 4096 | Chat, Summarize |
command-nightly | To reduce the time between major releases, we put out nightly versions of command models. For command, that is command-nightly.Be advised that command-nightly is the latest, most experimental, and (possibly) unstable version of its default counterpart. Nightly releases are updated regularly, without warning, and are not recommended for production use. | 8192 | Chat |
command-r | Command R is an instruction-following conversational model that performs language tasks at a higher quality, more reliably, and with a longer context than previous models. It can be used for complex workflows like code generation, retrieval augmented generation (RAG), tool use, and agents. | 128k | Chat |
command-r-plus | Command R+ is an instruction-following conversational model that performs language tasks at a higher quality, more reliably, and with a longer context than previous models. It is best suited for complex RAG workflows and multi-step tool use. | 128k | Chat |
Embed
These models can be used to generate embeddings from text or classify it based on various parameters. Embeddings can be used for estimating semantic similarity between two sentences, choosing a sentence which is most likely to follow another sentence, or categorizing user feedback, while outputs from the Classify endpoint can be used for any classification or analysis task. The Representation model comes with a variety of helper functions, such as for detecting the language of an input.
| Latest Model | Description | Dimensions | Max Tokens | Similarity Metric | Endpoints |
|---|---|---|---|---|---|
embed-english-v2.0 | Our older embeddings model that allows for text to be classified or turned into embeddings. English only | 4096 | 512 | Cosine Similarity | Classify, Embed |
embed-english-light-v2.0 | A smaller, faster version of embed-english-v2.0. Almost as capable, but a lot faster. English only. | 1024 | 512 | Cosine Similarity | Classify, Embed |
embed-multilingual-v2.0 | Provides multilingual classification and embedding support. See supported languages here. | 768 | 256 | Dot Product Similarity | Classify, Embed |
embed-english-v3.0 | A model that allows for text to be classified or turned into embeddings. English only. | 1024 | 512 | Cosine Similarity | Embed, Embed Jobs |
embed-english-light-v3.0 | A smaller, faster version of embed-english-v3.0. Almost as capable, but a lot faster. English only. | 384 | 512 | Cosine Similarity | Embed, Embed Jobs |
embed-multilingual-v3.0 | Provides multilingual classification and embedding support. See supported languages here. | 1024 | 512 | Cosine Similarity | Embed, Embed Jobs |
embed-multilingual-light-v3.0 | A smaller, faster version of embed-multilingual-v3.0. Almost as capable, but a lot faster. Supports multiple languages. | 384 | 512 | Cosine Similarity | Embed, Embed Jobs |
In this table we've listed older v2.0 models alongside the newer v3.0 models, but we recommend you use the v3.0 versions.
Rerank
The Rerank model can improve created models by re-organizing their results based on certain parameters. This can be used to improve search algorithms.
Rerank accepts full strings and than tokens, so the token limit works a little differently. Rerank will automatically chunk documents longer than 510 tokens, and there is therefore no explicit limit to how long a document can be when using rerank. See our best practice guide for more info about formatting documents for the Rerank endpoint.
Updated 22 days ago