Cohere Command

As AI consultants based in Rosenheim, Germany, we recommend Cohere Command for enterprises in the DACH region (Germany, Austria, Switzerland) focusing on RAG, enterprise search and agentic AI. Cohere is a Canadian company focused on enterprise LLMs; the Command family offers Command A+, Command A, Command A Reasoning, Command A Vision and Command R+ for knowledge-based applications. The models are available via the Cohere API, Microsoft Foundry and Hugging Face; fully on-premise deployment is also possible.

The currently relevant models are Command A+ (sparse MoE with 218B total and 25B active parameters, Apache 2.0, released 20 May 2026) for agentic workflows and multilingual applications, Command A for existing production workloads, and the embedding and rerank models (Embed 5, Rerank 5) for RAG. Cohere and Aleph Alpha are merging; details at the end of this page. This overview reflects the state as of October 2026.

Current models at a glance

Command A+

Command A+ is the first mixture-of-experts model of the Command family and its first frontier model under Apache 2.0:

• Sparse MoE: 218B total parameters, 25B active per token
• Unified capabilities: vision, reasoning, translation and agentic tool use in one model
• Efficiency: according to Cohere, runs on 1×B200 or 2×H100 thanks to W4A4 quantization, with 150% higher throughput than Command R+ 08-2024
• 48 languages including all official EU languages
• Native citations with span-level accuracy
• 128k context window (128k input, 64k output via the standard API)

Command A+ is positioned for sovereign critical infrastructure and, according to the changelog, is available to all users via the standard API.

Command A

Command A (command-a-03-2025) has been a dependable foundation for enterprise applications since March 2025 and remains a production-stable option: multi-step reasoning and autonomous task execution for complex workflows, plus translation in 23 languages. Command A Translate (August 2025) is a variant for multilingual workflows.

Command A Reasoning, Vision and Aya Vision

Command A Reasoning targets tasks requiring deep logical reasoning, with transparent chain-of-thought processing. Command A Vision adds image understanding: document OCR, image analysis and multimodal RAG workflows. The open-weights model Aya Vision (March 2025, Cohere Labs) provides multilingual image understanding under a CC-BY-NC licence, so it is mainly suited to research.

Command R+

Command R+ (command-r-plus-08-2024) with 104B parameters was long the first choice for RAG-intensive applications: higher throughput, lower latency, native citations with span-level accuracy, improved structured outputs and a 128k context window (about 320 pages). For new projects we recommend Command A or Command A+.

Cohere North

Cohere North is the enterprise platform for secure AI workspaces, available since January 2025 and used by companies such as RBC, Dell and Bell Canada. According to Cohere’s model documentation, “North” now also names a family of purpose-built models: North Small Translate (218B total / 25B active parameters, MoE, 50+ languages, CC BY-NC 4.0 on Hugging Face) and North Mini Code (30B total / 3B active parameters, MoE, for agentic coding), both available via the Cohere API.

RAG and enterprise search

For Cohere, RAG is not an add-on but a core competence. The combination of Command models, Embed 5 (Pro and Fast tiers, shared embedding space, Matryoshka dimensions 256 to 2048, multimodal inputs, 128k context, per the Cohere blog since 30 September 2026) and Rerank 5 (Pro and Fast, since 21 October 2026, with ‘Context Focus’ for the relevant passages in each document, calibrated relevance scores and 32k context) enables:

• Grounded generation – answers based on your documents
• Native citations – every statement with paragraph-level source attribution
• Precise reranking – optimised relevance ordering
• 128k context window – processing of extensive documents

The predecessors Embed v4 (multimodal for text, images and mixed documents such as PDFs) and Rerank v4.0 remain available. For preparing PDFs and forms, Cohere has also offered Parse (parse-v5.0) since 27 August 2026, which converts documents into structured Markdown.

EU availability and GDPR

Cohere Command can be run in the EU with several hosting options that differ by cloud provider and model:

• Microsoft Foundry: Command A (GA) and Command A+ (preview) can be selected with the “Global Standard” deployment type in twelve European regions according to Microsoft’s region documentation (including France Central, Germany West Central, Italy North, Spain Central, Sweden Central, West Europe; plus Norway, Switzerland and UK South and UK West in the United Kingdom). “Global Standard”, however, does not guarantee processing within the selected region. Microsoft also offers Rerank v4.0 in the “Data Zone Standard” deployment type in the EU data zone.
• AWS Bedrock: Command R and Command R+ were never available in Frankfurt, only in us-east-1 and us-west-2, and reached end of life on 19 August 2026 according to the AWS model card. Command A/A+ are not listed on Bedrock according to Cohere’s documentation. For Embed v4, Bedrock offers an EU geo cross-region profile across Frankfurt, Ireland, Paris, Milan, Spain and Stockholm, and Rerank 3.5 runs in-region in Frankfurt.
• Google Cloud: no Cohere models are listed in the model overview of the Gemini Enterprise Agent Platform (formerly Vertex AI), and Cohere does not name Google Cloud among its platforms.
• On-premise: full deployment for maximum data sovereignty; Command A+ can be self-hosted as an Apache 2.0 model.

For current Command models, Microsoft Foundry offers the widest choice of EU regions. For DACH companies with strict data protection requirements, on-premise remains the most reliable option.

Model lifecycle

• Command R and Command R+ (incl. command-r-plus-08-2024): reached end of life on Amazon Bedrock on 19 August 2026.
• Command A (March 2025): still usable in production; Command A+ is the successor for new projects.
• command-r-03-2024, command-r-plus-04-2024, command and command-light: deprecated since 15 September 2025 per Cohere. On the same date, Cohere discontinued fine-tuning via the dashboard and API; it offers custom model adaptation to enterprise customers on request. The versions command-r-08-2024 and command-r-plus-08-2024 remain available on the Cohere API.
• Aya Expanse 8B, Aya Vision 8B and Embed v2.0: no longer reachable via the Cohere API since 4 April 2026; the 32B variants of Aya remain available.

Merger with Aleph Alpha

On 24 April 2026, Cohere and Aleph Alpha (Heidelberg) announced the planned merger; on 16 September 2026 they signed the definitive merger agreement, according to the joint announcement. Key points: the combined company will operate worldwide as Cohere, with dual headquarters in Berlin and Toronto; Heidelberg remains a research centre; more than 1,000 employees combined. At closing, Aleph Alpha CEO Ilhan Scheer becomes Chief Operating Officer and co-founder Samuel Weinbach Chief Research Officer; Aidan Gomez remains CEO. Per the joint press release, the Schwarz Group has committed structured financing and is to lead Cohere’s upcoming Series E round. The sovereign offering for governments and regulated industries is to be provided on STACKIT. Further integration and product details have been announced; the announcements do not yet name a joint model or a timeline. Closing is expected later in 2026 per Cohere, subject to final regulatory approvals.

Independently of the merger, Aleph Alpha Research released Kolibri-1 on 3 October 2026, its own German-English MoE model under Apache 2.0. Neither company has yet communicated how Kolibri-1 and the Command family relate in the combined portfolio. Until closing, both companies remain independent; existing customers should clarify contractual questions directly with the respective vendor. More on our Aleph Alpha page.

Integration with CompanyGPT

Cohere Command is fully integrated into CompanyGPT. Use Cohere’s RAG strengths to make your internal knowledge searchable – GDPR-compliant in German or European data centres, or fully on-premise.

Our recommendation

Command A+ is the first choice for agentic workflows, tool use, multimodal and multilingual applications – especially for organisations that need an Apache 2.0 licence, sovereign deployment and EU language coverage. Command A remains the robust choice for existing production workloads. For RAG we recommend combining a Command model with Embed 5 and Rerank 5. For cost-efficient scenarios with simpler requirements, the compact Command R7B or Command R (08-2024) are suitable.

Note the weaknesses: Cohere is weaker at code generation than specialised models. For coding use cases we recommend OpenAI GPT or Anthropic Claude.