What it is
Command is the family of AI models made by Cohere, a company based in Toronto, Canada. Like GPT, Claude, and Gemini, a Command model is a language model: software trained on huge amounts of text so it can answer questions, summarise documents, draft emails, and follow instructions.
The difference is who it is for. Most people never meet Command directly, because Cohere has chosen not to compete for everyday consumers with a ChatGPT-style app. It sells to businesses and governments, especially in regulated fields such as banking, healthcare, energy, and the public sector, where data can't simply be sent to someone else's computers. A January 2025 CNBC interview with co-founder Nick Frosst presented Cohere as all about business, and happy to stay out of the consumer race led by ChatGPT. In an April 2024 Fortune interview, chief executive Aidan Gomez argued that skipping free consumer users also spares Cohere the heavy cost of running models for millions of people. That is the company's own account of its strategy; it is also a bet that business customers will stay loyal to a smaller supplier.
Who is behind it
Cohere was founded in 2019 by three founders who met through the University of Toronto: Aidan Gomez, Ivan Zhang, and Nick Frosst. Gomez, the chief executive, was one of the eight authors of "Attention Is All You Need", the 2017 Google paper that introduced the transformer, the design behind nearly every modern language model. He worked on it at age 20 as an intern at Google Brain. Frosst had worked at Google Brain in Toronto after studying under Geoffrey Hinton, and is also the lead singer of the indie band Good Kid.
The generations at a glance
Dates come from Cohere's own announcements and documentation. "B" means billions of parameters, the learned numbers inside a model.
- Command (November 2022). A beta model, then called command-xlarge, trained to respond well to single instructions.
- Command R (11 March 2024). Cohere built it for RAG, where a model answers from documents you supply, and for tool use, where it calls other software. It read up to 128,000 tokens at once and worked in 10 key languages.
- Command R+ (4 April 2024). A larger 104B version, launched first on Microsoft Azure, with answers that cite the passages they came from and support for multi-step tool use. Updated versions of both R models followed in August 2024.
- Command A (13 March 2025). A 111B model with a 256,000-token context window and 23 languages, which Cohere said could run on just two high-end Nvidia graphics cards (A100s or H100s) where comparable models needed far more.
- Command A Vision (July 2025) read images such as charts and scanned documents; Command A Reasoning and Command A Translate followed in August 2025.
- Command A+ (20 May 2026). A 218B mixture of experts model that uses about 25B parameters per word, supports 48 languages, reads text and images, and, according to Cohere, can run on two Nvidia H100s. Cohere's documentation calls it the last model in the Command A family.
At the time of writing Cohere also sells other models under new names, including North Mini Code for programming and Cohere Transcribe for speech. Model names and line-ups change often, so check Cohere's current model list.
The retrieval side: Embed and Rerank
Command is only half of what Cohere sells. Its Embed models turn text, and since Embed 4 (15 April 2025) images and scanned pages too, into embeddings: lists of numbers that capture meaning, so a search system can find passages about "staff holidays" even when the document says "annual leave". Those numbers are usually stored in a vector database. Its Rerank models then take the top search results and put them in order of relevance; Rerank 3.5 arrived in December 2024 and Rerank 4 in December 2025. Together they form the retrieval part of RAG, and many companies use Embed or Rerank with a different company's chat model.
North: the workplace app
In January 2025 Cohere opened early access to North, a workplace platform where staff can chat with the company's own records, search across tools such as email and Slack, write reports, and hand routine multi-step jobs to AI agents. It became generally available on 6 August 2025. Cohere named Royal Bank of Canada and Dell among its users, and TechCrunch listed LG, Ensemble Health Partners, and Palantir among companies that had piloted it.
Private deployment, and what the licences allow
Cohere's main selling point is where the model runs. Besides its own online service and big cloud marketplaces such as Microsoft Azure, Amazon, and Oracle, Cohere will install models and North inside a customer's own private cloud, on its own servers, or on machines cut off from the internet entirely (called "air-gapped"). Frosst told TechCrunch that Cohere can deploy "on a GPU in a closet". Data never has to leave the building, which matters for privacy rules and for governments wary of relying on foreign companies. Cohere now calls this "sovereign AI".
Cohere has also published model weights for anyone to download, with an important catch. From Command R in 2024 through Command A in 2025, research releases on Hugging Face came under the CC-BY-NC licence: free to study and experiment with, but "NC" means non-commercial, so a business needed a paid agreement with Cohere to use them in its products. That changed on 20 May 2026, when Cohere released Command A+ under Apache 2.0, a standard open-source licence that allows commercial use with attribution. Always check the licence on the model card for the exact version you download. Parameters explains what running a model yourself involves.
Money and partners
These figures come from the companies' announcements and news reports; valuations are set by investors and can change.
- June 2023: a 270 million US dollar funding round led by Inovia Capital, with Nvidia, Oracle, and Salesforce Ventures among the investors. The same month, Oracle announced it would build generative AI services with Cohere on its cloud and in its business software.
- December 2024: the Canadian government announced up to 240 million Canadian dollars for Cohere to buy computing capacity at a new data centre in Canada.
- August 2025: 500 million US dollars at a 6.8 billion dollar valuation, led by Radical Ventures and Inovia Capital. Joelle Pineau, formerly head of Meta's FAIR research lab, joined as chief AI officer.
- February 2026: CNBC reported, citing an investor memo, that Cohere reached about 240 million dollars in annual recurring revenue in 2025, beating its 200 million target.
- April to September 2026: Cohere agreed to combine with Aleph Alpha, a German AI company, in a deal TechCrunch reported valued the combined firm at about 20 billion dollars, with backing from Germany's Schwarz Group, owner of Lidl. A definitive agreement was signed on 16 September 2026. At the time of writing it still needs regulatory approval and is expected to close later in 2026, with the merged company keeping the Cohere name and headquarters in Toronto and Berlin.
Why it matters
It shows another way to run an AI company. OpenAI, Google, and Anthropic each fund famous consumer apps. Cohere has so far built a business mainly on selling to other businesses. Whether that approach can keep up with far richer rivals is an open question, and worth watching.
It pushed RAG and citations into the mainstream. Command R was one of the first major models marketed mainly for answering from a company's own documents, with citations, rather than for general chat. Patrick Lewis, lead author of the paper that named RAG, went on to lead RAG work at Cohere, as told on the RAG page. Answers that point to their sources are easier to check, which helps with hallucination and grounding.
It keeps AI in-house for cautious organisations. For a hospital, a bank, or a defence ministry, "send your data to our cloud" can be a dealbreaker. Models that run on two graphics cards behind the customer's own firewall make AI possible in places it otherwise wouldn't go.
It is part of a wider push for AI outside the United States. With Canadian government backing, the Aleph Alpha deal, and a focus on many languages, Cohere pitches itself to governments that want capable AI without depending on American tech giants. Critics note that funding from US investors and a possible stock market listing could complicate that promise, a point TechCrunch raised in April 2026.
Real-world examples
The publishers' lawsuit (February 2025 to the present)
On 13 February 2025, fourteen news and magazine publishers sued Cohere in federal court in New York. The case is called Advance Local Media v. Cohere, and the plaintiffs include Condé Nast, The Atlantic, Forbes, The Guardian, Business Insider, the Los Angeles Times, Politico, the Toronto Star, and Vox Media. They alleged that Cohere copied at least 4,000 of their articles to train its models, including from behind paywalls, and that Command reproduced large parts of articles, sometimes word for word, when users asked about the news. They also claimed that when Command hallucinated articles that never existed and attributed them to real publications, it infringed their trademarks.
Cohere's head of communications, Josh Gartner, said the company "strongly stands by its practices for responsibly training its enterprise AI" and called the lawsuit "misguided and frivolous".
On 13 November 2025, Judge Colleen McMahon refused Cohere's request to dismiss parts of the case. She ruled that the publishers had made plausible claims, including that AI summaries close enough to replace an article could infringe copyright, and that invented articles carrying a publisher's name could be a trademark problem. This was not a finding that Cohere broke the law; it only allowed the case to go forward. At the time of writing, court records show the case in the evidence-gathering stage, with no trial or settlement reported. Similar cases against other AI companies are covered on the Datasets page.
Aya: 101 languages from a volunteer army (February 2024)
Most AI models work best in English and a handful of other widely used languages. On 13 February 2024, Cohere's non-profit research arm, then called Cohere For AI (renamed Cohere Labs in 2025), released Aya 101, a 13B model that follows instructions in 101 languages, more than 50 of them poorly served by earlier models. It came with a large training dataset, much of it written or checked by more than 3,000 volunteer researchers and speakers in 119 countries, who contributed examples in their own languages. Unlike the Command research releases, Aya 101 was published under the permissive Apache 2.0 licence.
The project continued with Aya 23 (May 2024), Aya Expanse, Aya Vision, and in February 2026 Tiny Aya, small models meant to run on ordinary devices. The research page lists awards including an ACL 2024 best paper award. Aya matters beyond Cohere: it showed that people outside big labs could help build AI for their own languages, a theme also covered in Translation.
An everyday case: the insurance claims team
Picture Marta, who runs a claims team at a mid-sized insurance company. Her company's rules forbid pasting customer files into a public chatbot. The IT department sets up a Command model inside the company's private cloud, connected to the policy handbook and past claim decisions through Embed and Rerank. When a claims handler asks "Is flood damage to a garden shed covered under the standard home policy?", the system finds the relevant handbook sections and answers with numbered citations. Marta's rule for her team: click every citation before replying to a customer, because the model can still choose the wrong passage or misread a clause, and the final decision stays with a person, as described in Human in the Loop.
Try it yourself
This exercise tests Command's best-known strength, answering only from text you give it and showing where each answer came from. It uses Cohere's free demo of Command A+ on Hugging Face. At the time of writing it opened in a web browser without an account or payment; availability can change.
- Open the demo. Go to huggingface.co/spaces/CohereLabs/command-a-plus-05-2026. The chat box sits under the heading "Command A+". The page notes that by using it you agree to Cohere's privacy policy, so type nothing private or personal.
- Give it a short document and a question. Copy and paste this: > Using only the text below, answer the question. After each sentence of your answer, quote the sentence from the text that supports it. If the text does not contain the answer, say "Not in the text." > > Text: "Riverside Library opens at 9 am on weekdays and 10 am on Saturdays. It is closed on Sundays and public holidays. Members may borrow up to 12 items for three weeks. Laptops can be borrowed for one day only, and only by members over 16." > > Question: Can a 15-year-old member borrow a laptop on Saturday, and when is the library open that day?
- Check the quotes. Compare each quoted sentence with the text. Does every claim have a matching quote? Did it combine the age rule and the opening hours correctly?
- Ask something the text doesn't cover. Paste this: > Using only the same text, how much is the fine for a late book? A well-grounded answer says the text doesn't say. If it invents a fine, you have seen a hallucination at work.
- Try it in another language. Command A+ is listed as supporting 48 languages. Paste: > Translate your first answer into Spanish and then into Japanese, keeping the quoted sentences in English. If you speak one of those languages, check the result. Ask someone who does if you don't.
- Reflect. In a real business system the "text" would be retrieved automatically from thousands of documents, as described in RAG. The quote-checking habit you just practised is what every user of such a system should keep.
If the demo is unavailable: Cohere offers a free, rate-limited trial key when you sign up at dashboard.cohere.com; at the time of writing it allowed 1,000 calls a month and included a playground. You can also paste the same prompts into any chatbot to compare how well each one sticks to the text.
Common mistakes and misunderstandings
Thinking Cohere has no public product at all. It has no mass-market consumer app, but developers can sign up for a trial, and Cohere Labs publishes demos and research models. What it sells is aimed at organisations.
Assuming citations mean the answer is right. Citations let you check an answer; they don't guarantee it. A model can cite the right document and still summarise it wrongly, or cite a passage that doesn't quite say what it claims. Open the source.
Treating "downloadable" as "free for business". Most Command research releases before May 2026 were CC-BY-NC: fine for experiments, not for commercial products without a deal with Cohere. Command A+ is Apache 2.0. The licence depends on the exact model, so read the model card.
Mixing up the family names. Command models generate text. Embed and Rerank models power search and don't write answers. North is an app built on top of them. Aya is Cohere Labs' multilingual research family. A company "using Cohere" might be using any one of these.
Believing private deployment solves every risk. Running a model on your own servers keeps your data in-house, but it doesn't stop wrong answers, and you become responsible for security, updates, and access control yourself.
Relying on benchmark claims in launch posts. Cohere, like every company, compares its models to rivals on tests it chooses. Treat those claims as the company's own and test the model on your real tasks. Benchmarks explains why.
Forgetting that things are moving fast. Model names, licences, trial limits, and even the company's structure (with the Aleph Alpha merger pending at the time of writing) change quickly. Check Cohere's current documentation before making decisions.
Related elements
- RAG: the technique Command R was built around, and the story of how RAG got its name.
- Embeddings: what Cohere's Embed models produce and why they make search smarter.
- Vector Database: where embeddings are stored so the right passages can be found fast.
- Grounding: tying answers to trusted sources, which Command's citations are designed to support.
- Tool Use: how Command models call other software to get work done.
- Privacy: why keeping a model on your own servers matters to hospitals, banks, and governments.
- Transformer: the design Aidan Gomez helped introduce in 2017.
- Translation: where multilingual work like Aya fits in.
- Parameters: what it takes to run a downloadable model on your own hardware.
- Mistral: another company outside the United States selling AI on the promise of independence and control.