Command (Cohere)
Overview
Cohere is an enterprise-focused AI company. Founded in 2019 in Toronto, Cohere has carved a unique position as an LLM provider specializing in enterprise RAG, retrieval, and agents rather than consumer chat. The current top model is Command A+, open-sourced under Apache 2.0 on May 19, 2026: a 218B sparse MoE that folds the four earlier Command A variants into one, with 256K context, 48 languages, and Cohere's first multimodal reasoning. Quantized to W4A4 it runs on as few as two H100s.
Current model: Command A+
Cohere's top model, specialized for RAG, tool use, agents, and multilingual work.
Performance and efficiency
| Metric | Value |
|---|---|
| Architecture | 218B sparse MoE (consolidating the four Command A variants) |
| GPU requirement | From 2 H100s when quantized to W4A4 |
| Context length | 256,000 tokens (~600 pages) |
| Languages | 48 languages |
| License | Apache 2.0 (open source) |
Key Features
1. Efficient GPU Requirements
Runs on just 2 A100/H100 GPUs — far fewer than the 32 required by some competitors. Dramatically lowers private deployment costs.
2. Enterprise RAG Specialization
Optimized for high-accuracy retrieval-augmented generation. Produces verifiable answers with citations.
3. Agentic Tool Use
Supports complex workflows with agent-style tool use.
4. North AI Platform
Cohere's enterprise agent platform for secure in-house automation.
5. Model Vault (Launched September 2025)
Dedicated inference platform. Deploy Command, Rerank, and Embed in isolated VPCs or on-premises — sensitive data never leaves your organization.
6. Multimodal support
Command A+ is Cohere's first multimodal reasoning model, folding enterprise-grade image understanding into the main API and superseding the vision-only Command A Vision.
7. North Mini Code 1.0 (open-source coding model)
Cohere's first developer-specific model: a 30B MoE that activates 3B parameters per token and runs on a single NVIDIA H100. The weights are public.
Specialized APIs (Key to RAG)
Rerank API
Specialized for reranking search results. Significantly improves RAG pipeline accuracy.
Embed API (Multilingual)
Multilingual embedding model. The current version is Embed v4 at $0.10 per 1M tokens. Cohere is recognized as industry-leading in multilingual RAG.
Pricing
| Item | Price |
|---|---|
| Command A+ / Command A API | $2.50 / $10 per 1M tokens |
| Command R API | $0.15 / $0.60 per 1M tokens |
| Command R7B API | $0.0375 / $0.15 per 1M tokens |
| Embed v4 | $0.10 per 1M tokens |
| Rerank | $2.00 per 1,000 searches |
| Private deployment | Up to 50% cheaper than API |
Strengths
- ✅ High-accuracy enterprise RAG
- ✅ 48-language multilingual processing and embeddings
- ✅ Command A+ is Apache 2.0 open source and self-hostable from two H100s
- ✅ Model Vault for on-prem / VPC deployment
- ✅ North AI platform for agentic automation
- ✅ North Mini Code 1.0 covers coding workloads too
Weaknesses
- ❌ No consumer product (no ChatGPT/Claude.ai-like app)
- ❌ Lower brand awareness than OpenAI/Anthropic
- ❌ No image/video generation
- ❌ No consumer free tier
Official Resources
- Website: https://cohere.com
- API Docs: https://docs.cohere.com
- Command A: https://docs.cohere.com/docs/command-a
- North AI: https://cohere.com/north