# Cohere

> Source: https://aiwiki.ai/wiki/cohere
> Summary: Cohere is a Canadian artificial intelligence company that develops language, retrieval, speech, and multimodal models for businesses and public-sector organizations. It was founded in Toronto in 2019 by Aidan Gomez, Nick Frosst, and Ivan Zhang.
> Updated: 2026-07-31
> Fact-checked: 2026-07-31
> Categories: AI Companies, Large Language Models, Natural Language Processing
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Cohere." aiwiki.ai, 31 Jul 2026. https://aiwiki.ai/wiki/cohere
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Cohere | |
| --- | --- |
| **Type** | Privately held company |
| **Industry** | Artificial intelligence |
| **Founded** | 2019 |
| **Founders** | Aidan Gomez, Nick Frosst, Ivan Zhang |
| **Headquarters** | Toronto, Ontario, Canada |
| **Key people** | Aidan Gomez (co-founder and CEO), Phil Blunsom (CTO), Joelle Pineau (chief AI officer), François Chadwick (CFO) |
| **Products** | Command, Embed, Rerank, North, Compass, Model Vault |
| **Website** | [cohere.com](https://cohere.com/) |

Cohere is a Canadian [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company that develops language, retrieval, speech, and multimodal models for businesses and public-sector organizations. It was founded in Toronto in 2019 by [Aidan Gomez](https://aiwiki.ai/wiki/aidan_gomez), Nick Frosst, and Ivan Zhang. Its commercial model families include Command for generation and agentic tasks, Embed for vector representations, and Rerank for relevance scoring. Cohere also sells North, an enterprise agent platform, and Compass, an enterprise search system. Customers can use some Cohere models through a managed API or cloud marketplace, while Cohere also offers private deployments in a virtual private cloud or on customer-controlled infrastructure.[1][25]

The company has consistently emphasized enterprise use rather than operating a general-purpose consumer chatbot. That positioning affects its model design, which stresses retrieval, multilingual operation, tool use, and deployment on a comparatively small number of accelerators. It also explains the importance of private deployment and data residency in Cohere's product line. Performance and efficiency figures published by Cohere are vendor evaluations unless an independent evaluation is identified.[12][13][25]

## History

### Founding

Aidan Gomez was one of the eight authors of the 2017 paper "[Attention Is All You Need](https://aiwiki.ai/wiki/attention_is_all_you_need)," which introduced the [Transformer](https://aiwiki.ai/wiki/transformers) architecture. The paper replaced recurrent and convolutional sequence processing with an architecture based primarily on attention, and it became foundational to later [large language models](https://aiwiki.ai/wiki/large_language_model).[2] Gomez, Nick Frosst, and Ivan Zhang were all former University of Toronto students.[3]

Gomez and Zhang had also established the volunteer research community for.ai in 2017. Cohere's account of its research history says that this community was brought into the company in June 2022 as Cohere For AI, then renamed Cohere Labs in April 2025.[4] This research organization is distinct from the commercial product group, although both operate within Cohere.

The founders established Cohere in 2019. Early work focused on making language models available through developer interfaces and on improving multilingual text representation. As the generative AI market expanded, Cohere concentrated on models and systems intended for enterprise search, retrieval-augmented generation, and controlled deployment.[1][3]

### Financing and growth

Cohere announced a US$270 million Series C financing in June 2023. The round was led by Inovia Capital, with participation from investors including NVIDIA, Oracle, Salesforce Ventures, DTCP, Mirae Asset, Schroders Capital, SentinelOne, and Thomvest Ventures.[5] In July 2024, the company raised another US$500 million at a reported US$5.5 billion valuation. TechCrunch reported that the 2024 round brought Cohere's total raised at that point to about US$970 million.[6]

Shortly after the July 2024 financing, Cohere eliminated about 20 positions, then roughly 5 percent of its staff. The company said it would continue hiring in selected areas. The timing attracted attention because the reduction followed immediately after the funding announcement, but the event did not represent a company-wide closure or withdrawal from model development.[7]

In August 2025, Cohere announced a US$500 million round at a US$6.8 billion valuation. Radical Ventures and Inovia Capital led the round, with participation from several existing and strategic investors.[8] An additional US$100 million extension was reported the following month, taking the reported valuation to about US$7 billion.[9] These financing announcements should not be combined mechanically with government project support when stating how much equity capital the company has raised.

The Government of Canada finalized a contribution of up to C$240 million in March 2025 toward a C$725 million Cohere project involving domestic computing capacity and the scaling of AI capabilities. The announcement described a public contribution to a defined project, not an unrestricted venture-capital round.[10]

Private-company revenue figures are less transparent. In February 2026, TechCrunch, citing an investor memorandum previously reported by CNBC, wrote that Cohere reached US$240 million in annual recurring revenue for 2025 and exceeded a US$200 million internal target.[11] This is a reported ARR figure, not an audited public-company revenue statement.

| Event | Date | Amount or reported valuation | Evidentiary status |
| --- | --- | --- | --- |
| Series C | June 2023 | US$270 million | Announced by Cohere |
| Financing round | July 2024 | US$500 million at a reported US$5.5 billion valuation | Independently reported |
| Financing round | August 2025 | US$500 million at a US$6.8 billion valuation | Announced by Cohere |
| Round extension | September 2025 | US$100 million; about US$7 billion valuation | Independently reported |
| Canadian compute project contribution | March 2025 | Up to C$240 million toward a C$725 million project | Government announcement |
| Proposed Aleph Alpha combination and Series E | April 2026 | €500 million, described as about US$600 million, intended from Schwarz Group | Announced plan, subject to approvals at announcement |

### Leadership and locations

Gomez remains Cohere's chief executive. As of July 28, 2026, Cohere's leadership page listed Phil Blunsom as chief technology officer, Joelle Pineau as chief AI officer, and François Chadwick as chief financial officer.[1] Pineau, previously a senior AI research leader at Meta and a professor at McGill University, joined Cohere in 2025.[39]

Cohere identifies Toronto as its home and lists offices in Toronto, Montreal, New York, San Francisco, London, Paris, and Seoul.[1] The company announced a larger London office in June 2026, framing the expansion around research, engineering, and customer work in the United Kingdom.[40] Office lists change over time and do not establish a separate legal headquarters in every listed city.

## Models

Cohere's models fall into several functional groups. Command models generate or transform content and can support tools and agents. Embed models map content into vectors for retrieval and classification. Rerank models score candidate documents against a query after an initial search. Newer releases also cover speech recognition and coding agents.[13][17][18][19][20]

### Command

[Command A](https://aiwiki.ai/wiki/command_a), released in March 2025, is a 111-billion-parameter dense Transformer with a 256,000-token context window. Cohere's technical report describes post-training for [retrieval-augmented generation](https://aiwiki.ai/wiki/retrieval_augmented_generation), [tool use](https://aiwiki.ai/wiki/tool_use), multilingual interaction, and enterprise tasks. The report and product documentation say that a quantized deployment can run on two A100 or [H100](https://aiwiki.ai/wiki/nvidia_h100) GPUs and supports 23 languages.[12][13] That is a deployment claim for Cohere's specified configuration, not a guarantee that every workload will achieve the same throughput or memory use.

The technical report describes Command A as a decoder-only Transformer that interleaves three sliding-window attention layers with one full-attention layer. Sliding-window layers use rotary position embeddings, while full-attention layers do not use positional embeddings. The architecture also uses grouped-query attention, SwiGLU activations, shared input and output embeddings, and no bias terms. Cohere post-trained separate specialists for code, safety, retrieval, mathematics, multilingual work, and long context before merging them and conducting a final preference-optimization stage.[12] These design details help explain the company's efficiency objective, although they do not independently establish superiority over another model.

Command A Vision, released in July 2025, added image input for document and visual understanding. Cohere's current documentation gives it a 128,000-token context window and support for up to 20 images in one request. The documentation also states that this model does not support tool use or image generation.[14]

Command A Reasoning followed in August 2025. It retains 111 billion parameters and a 256,000-token context window, with a maximum output of 32,000 tokens. Developers can enable or disable a reasoning mode and set a token budget for that process. Cohere describes it as supporting English and 22 additional languages and as being intended for complex tool-using and agentic tasks.[15]

In May 2026, Cohere released Command A+, a sparse mixture-of-experts model under the Apache 2.0 license. Cohere's detailed technical announcement specifies 218 billion total parameters with 25 billion active for a token, a 128,000-token input context, text and image input, and support for 48 languages. The company says a quantized version can run on two H100 GPUs.[16] Benchmark results in that announcement were produced or selected by Cohere and should be read as vendor-reported evaluations.

| Model | Release | Architecture or size | Context | Notable scope |
| --- | --- | --- | --- | --- |
| Command A | March 2025 | Dense, 111B parameters | 256K | Text generation, RAG, tools, 23 languages |
| Command A Vision | July 2025 | Multimodal Command variant | 128K | Text and image input, no tool use or image generation |
| Command A Reasoning | August 2025 | 111B parameters | 256K | Optional reasoning mode and tool-oriented tasks |
| Command A+ | May 2026 | Sparse MoE, 218B total and 25B active | 128K input | Multimodal reasoning, 48 languages, Apache 2.0 |

Older [Command R](https://aiwiki.ai/wiki/command_r) and Command R+ models remain relevant because they were designed specifically for retrieval-augmented generation and tool use. They also became available through third-party platforms such as Amazon Bedrock.[31] Hosted-model prices and marketplace availability can change independently of a model's technical identity.[25][31]

### Embed and Rerank

Embed 4, announced in April 2025, creates shared [embeddings](https://aiwiki.ai/wiki/embeddings) for text, images, and mixed documents. Cohere documents a context length of up to 128,000 tokens and output dimensions of 256, 512, 1,024, or 1,536 through Matryoshka representation learning. Its intended applications include semantic search, classification, clustering, and retrieval over visually rich documents.[17]

A typical Cohere retrieval pipeline first uses Embed to find a broad candidate set, then applies Rerank to order those candidates for a specific query. This is a two-stage information retrieval design, not a guarantee that generation will be factually correct. Retrieval quality still depends on the indexed corpus, chunking, access controls, query formulation, and evaluation method.[17][18]

Rerank 4.0, announced in December 2025, is offered in Pro and Fast variants. Cohere documents a 32,000-token context and support for text as well as semi-structured content such as JSON, tables, and code. A reranker returns relevance scores and a new ordering; it does not generate the final answer.[18]

| Family | Input and output | Role in a retrieval system |
| --- | --- | --- |
| Embed | Content to dense vector | Candidate retrieval, clustering, classification |
| Rerank | Query and candidate documents to relevance scores | More precise second-stage ordering |
| Command | Prompt and retrieved evidence to generated response | Synthesis, tools, and grounded generation |

### Speech and code models

Cohere Transcribe, released in March 2026, is a 2-billion-parameter automatic speech recognition model. Cohere describes a Conformer encoder and Transformer decoder, support for 14 languages, and release under Apache 2.0.[19] Claims that it led particular speech benchmarks at release are vendor claims and are not repeated here as independent findings.

North Mini Code, released in June 2026, is a coding model designed for agentic software tasks. Cohere describes it as a sparse mixture-of-experts model with 30 billion total parameters and 3 billion active parameters, a 256,000-token input context, and up to 64,000 output tokens. The weights are available under Apache 2.0.[20] Despite its name, it is a model that can be used within coding agents, not the North workplace application itself.

## Products and deployment

### North

North combines conversational assistance, enterprise search, tools, and workflow automation in an employee-facing platform. Cohere opened an early-access version in January 2025 and announced general availability in August 2025.[22] The product can connect to organizational data and applications, and Cohere offers it for managed, virtual private cloud, and on-premises deployment.[21]

North is distinct from the underlying Command, Embed, and Rerank models. The models provide generation and retrieval capabilities, while North packages those capabilities with connectors, identity and access features, agent construction, and an end-user workspace. Cohere's claims about security, time savings, and customer outcomes are product claims unless supported by a named independent assessment.[21][22]

### Compass and Model Vault

Compass is Cohere's end-to-end search and discovery system. It ingests supported enterprise files and data sources, parses and indexes their contents, and uses Embed and Rerank in a managed search pipeline. Cohere offers Compass in a virtual private cloud, on premises, or within Model Vault.[23]

Model Vault, announced in January 2026, is a deployment option in which Cohere manages models inside an isolated environment intended for a particular customer. It is meant to combine managed operation with stronger separation than a multi-tenant public API.[24] Whether a deployment meets a particular regulatory or security standard depends on the actual contract, architecture, controls, and audit scope, not solely on the Model Vault label.

### APIs, cloud services, and private deployment

Cohere documents several delivery paths: its own platform API, cloud AI services, virtual private cloud deployments, and on-premises installations. Listed cloud channels include [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services) Bedrock and SageMaker, Microsoft Azure AI Foundry, and [Oracle](https://aiwiki.ai/wiki/oracle) Cloud Infrastructure.[25] Availability varies by model, region, and provider. Amazon's documentation, for example, lists selected Command, Embed, and Rerank models rather than every Cohere release.[31]

Coral, introduced in 2023, was an earlier enterprise knowledge assistant that demonstrated grounded answers over connected data.[26] North later became Cohere's broader agent and workplace platform. Treating Coral, North, and Command as interchangeable obscures the distinction between an application, a platform, and a model family.

## Cohere Labs and Aya

Cohere Labs conducts open research and coordinates a distributed research community. One of its central projects is Aya, which investigates multilingual instruction tuning and evaluation beyond a small set of high-resource languages.[4]

The original Aya project produced an instruction-tuned model covering 101 languages. Its ACL 2024 paper describes a collaborative effort involving researchers from many countries and emphasizes that more than half of the languages represented were lower-resource languages.[27] The released model and dataset were research artifacts; their coverage does not imply equal quality across every language or task.

Aya Expanse, described in a 2024 technical report, consists of 8-billion and 32-billion-parameter multilingual models spanning 23 languages. The work combined multilingual data, preference optimization, and model merging methods.[28] Aya Vision extended the program to multilingual image and text understanding in 8-billion and 32-billion-parameter models.[29]

Tiny Aya, described in a March 2026 paper, reduced the model size to 3.35 billion parameters while targeting 70 languages. The project includes a global model and region-focused variants. The paper evaluates translation, language understanding, safety, and on-device efficiency, but its comparative results remain evaluations by the model's authors.[30]

| Aya release | Model sizes | Language or modality focus |
| --- | --- | --- |
| Aya 101 | 13B | Instruction tuning across 101 languages |
| Aya Expanse | 8B and 32B | Generative models across 23 languages |
| Aya Vision | 8B and 32B | Multilingual image and text understanding |
| Tiny Aya | 3.35B | 70 languages with global and regional variants |

## Partnerships and organizational changes

Cohere models are distributed through infrastructure partners as well as Cohere's own platform. Amazon documents Cohere models in Bedrock, and Cohere's deployment documentation also identifies Azure and Oracle channels.[25][31] Fujitsu developed Takane, a Japanese enterprise model based on Cohere's Command series, as part of a partnership with Cohere.[32] In July 2026, the University of Toronto and Cohere announced a partnership for institution-wide access, research, education, and evaluation of enterprise AI tools.[3]

### Proposed combination with Aleph Alpha

On April 24, 2026, Cohere and the German AI company Aleph Alpha announced an agreement to combine. The transaction was subject to regulatory and shareholder approvals. Cohere also said Schwarz Group intended to lead a Cohere Series E financing with €500 million, described in the announcement as approximately US$600 million.[33][34]

Independent reporting described the proposed combined company as having an implied value of about US$20 billion and said Cohere would lead the entity.[34] That figure was a transaction-related reported valuation, not evidence that Cohere had already completed a US$20 billion financing. Both Cohere's announcement and the independent report described a planned transaction rather than a completed merger.[33][34]

### Reliant AI

In May 2026, Cohere announced that it had acquired Reliant AI, a company developing research agents for scientific and technical work. Cohere said Reliant's team and technology would be incorporated into its sovereign and enterprise AI work.[35] Because the acquisition announcement came from Cohere, claims about the acquired technology's performance should be treated as the company's description.

## Copyright and trademark litigation

In February 2025, a group of publishers filed *Advance Local Media LLC et al. v. Cohere Inc.* in the US District Court for the Southern District of New York. The complaint alleges copyright infringement in model training and real-time retrieval, infringement through outputs, contributory infringement, and trademark violations. It also alleges that Cohere systems reproduced publisher material and sometimes attached publisher brands to inaccurate output.[36] These are allegations in a civil complaint, not adjudicated facts.

Cohere disputed the allegations and in May 2025 filed a partial motion to dismiss. Its public response said the company stood by its training practices and argued that portions of the complaint failed as a matter of law.[37] On November 13, 2025, the court denied the partial motion to dismiss, allowing the challenged claims to proceed.[38] A denial at the pleading stage is not a judgment that the publishers proved infringement. It means the challenged claims were not dismissed at that stage.

## See also

- [Anthropic](https://aiwiki.ai/wiki/anthropic)
- [Mistral AI](https://aiwiki.ai/wiki/mistral_ai)
- [OpenAI](https://aiwiki.ai/wiki/openai)

## References

1. Cohere. "About Cohere." https://cohere.com/about
2. Ashish Vaswani et al. "Attention Is All You Need." *Advances in Neural Information Processing Systems 30*, 2017. https://arxiv.org/abs/1706.03762
3. University of Toronto. "U of T partnership with Cohere sets stage for responsible AI adoption at scale." July 16, 2026. https://www.utoronto.ca/news/u-t-partnership-cohere-sets-stage-responsible-ai-adoption-scale
4. Cohere Labs. "Research." https://cohere.com/research
5. Cohere. "Cohere Raises $270M Series C to Bring Generative AI to the Enterprise." June 8, 2023. https://cohere.com/blog/announcement
6. Kyle Wiggers. "Cohere raises $500M to beat back generative AI rivals." *TechCrunch*, July 22, 2024. https://techcrunch.com/2024/07/22/cohere-raises-500m-to-beat-back-generative-ai-rivals/
7. Sheryl Estrada. "After $500 million funding, AI startup Cohere lays off employees." *Fortune*, July 23, 2024. https://fortune.com/2024/07/23/after-500-million-funding-ai-startup-cohere-layoffs/
8. Cohere. "Cohere raises $500M USD to accelerate the delivery of secure AI for global enterprises." August 2025. https://cohere.com/blog/august-2025-funding-round
9. Kyle Wiggers. "Cohere hits $7B valuation a month after its last raise, partners with AMD." *TechCrunch*, September 24, 2025. https://techcrunch.com/2025/09/24/cohere-hits-7b-valuation-a-month-after-its-last-raise-partners-with-amd/
10. Innovation, Science and Economic Development Canada. "Government of Canada finalizes investment to support Canadian-Born AI leader, Cohere." March 20, 2025. https://www.canada.ca/en/innovation-science-economic-development/news/2025/03/government-of-canada-finalizes-investment-to-support-canadian-born-ai-leader-cohere.html
11. Rebecca Bellan. "Cohere's $240M year sets stage for IPO." *TechCrunch*, February 13, 2026. https://techcrunch.com/2026/02/13/coheres-240m-year-sets-stage-for-ipo/
12. Cohere. "Command A: An Enterprise-Ready Large Language Model." Technical report, 2025. https://cohere.com/research/papers/command-a-technical-report.pdf
13. Cohere Documentation. "Introducing Command A." March 13, 2025. https://docs.cohere.com/changelog/command-a
14. Cohere Documentation. "Command A Vision." https://docs.cohere.com/docs/command-a-vision
15. Cohere Documentation. "Announcing Cohere's Command A Reasoning Model." August 21, 2025. https://docs.cohere.com/changelog/2025-08-21-command-a-reasoning
16. Cohere. "Introducing Command A+: Making sovereign agentic capabilities available to all." May 20, 2026. https://cohere.com/blog/command-a-plus
17. Cohere. "Introducing Embed 4: Multimodal Search for Business." April 15, 2025. https://cohere.com/blog/embed-4
18. Cohere Documentation. "Introducing Rerank 4.0." December 11, 2025. https://docs.cohere.com/changelog/rerank-v4.0
19. Cohere. "Introducing Cohere Transcribe." March 26, 2026. https://cohere.com/blog/transcribe
20. Cohere. "Introducing North Mini Code." June 9, 2026. https://cohere.com/blog/north-mini-code
21. Cohere. "North." https://cohere.com/north
22. Kyle Wiggers. "Cohere's new AI agent platform North promises to keep enterprise data secure." *TechCrunch*, August 6, 2025. https://techcrunch.com/2025/08/06/coheres-new-ai-agent-platform-north-promises-to-keep-enterprise-data-secure/
23. Cohere. "Compass: Intelligent Search and Discovery." https://cohere.com/compass
24. Cohere. "Introducing Model Vault." January 28, 2026. https://cohere.com/blog/model-vault
25. Cohere Documentation. "Deployment Options Overview." https://docs.cohere.com/docs/deployment-options-overview
26. Cohere. "Introducing Coral: The Knowledge Assistant for Enterprises." 2023. https://cohere.com/blog/introducing-coral-the-knowledge-assistant-for-enterprises
27. Viraat Aryabumi et al. "Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model." *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics*, 2024. https://aclanthology.org/2024.acl-long.845/
28. Cohere For AI. "Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontier." 2024. https://arxiv.org/abs/2412.04261
29. Cohere Labs. "Aya Vision: Advancing the Frontier of Multilingual Multimodality." 2025. https://arxiv.org/abs/2505.08751
30. Cohere Labs. "Tiny Aya: Open Multilingual Models for On-Device Intelligence." 2026. https://arxiv.org/abs/2603.11510
31. Amazon Web Services. "Cohere models in Amazon Bedrock." https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-cohere.html
32. Cohere. "Fujitsu launches Takane AI model in partnership with Cohere." May 20, 2025. https://cohere.com/customer-stories/fujitsu
33. Cohere. "Cohere and Aleph Alpha Join Forces." April 24, 2026. https://cohere.com/blog/cohere-alephalpha-join-forces
34. Rebecca Bellan. "Why Cohere is merging with Aleph Alpha." *TechCrunch*, April 25, 2026. https://techcrunch.com/2026/04/25/why-cohere-is-merging-with-aleph-alpha/
35. Cohere. "Cohere acquires Reliant AI to expand sovereign enterprise AI." May 19, 2026. https://cohere.com/blog/cohere-acquires-reliant-ai-expand-sovereign-enterprise-ai
36. *Advance Local Media LLC et al. v. Cohere Inc.*, Complaint, No. 1:25-cv-01305, US District Court for the Southern District of New York, February 13, 2025. https://business.cch.com/ipld/AdvanceLocalMediaCohereComp20250213.pdf
37. Cohere. "Cohere files motion to dismiss in Advance Local Media suit." May 23, 2025. https://cohere.com/blog/publisher-response
38. *Advance Local Media LLC et al. v. Cohere Inc.*, Opinion and Order, Document 59, No. 1:25-cv-01305, US District Court for the Southern District of New York, November 13, 2025. https://business.cch.com/ipld/AdvanceLocalMediaCohere20251113.pdf
39. Maxwell Zeff. "Cohere hires longtime Meta research head Joelle Pineau as its chief AI officer." *TechCrunch*, August 14, 2025. https://techcrunch.com/2025/08/14/cohere-hires-long-time-meta-research-head-joelle-pineau-as-its-chief-ai-officer/
40. Cohere. "Cohere triples UK footprint with new London office to support R&D growth." June 15, 2026. https://cohere.com/blog/cohere-triples-uk-footprint-with-new-london-office-to-support-r-and-d-growth