Skip to content

Cohere — The RAG-Native Outlier: How Command, Embed, Rerank, and Aya Fit Together

Aug 24, 2026 1 min
TL;DR Cohere is the only family that ships generation, retrieval, reranking, and multilingual as distinct products. Command A runs 256K context on two GPUs at 111B, Embed v4 does mixed image-text retrieval, Rerank v4 handles 32K semi-structured data, and Aya covers 101 languages — a four-piece stack built for RAG. This post breaks down each pillar's positioning, licensing, and selection guide.
Table of Contents
  1. Family Timeline
  2. Four Pillars: Why Cohere Is Not Just an LLM Company
  3. Architecture: Why "RAG-Native"
    1. Grounded generation with citations
    2. Multi-step tool use
    3. Efficiency-first design
  4. Embed v4 and Rerank v4: How to Choose
    1. Embed: From Dual Track to Unified
    2. Rerank: From 4K to 32K
    3. One-Line Selector
  5. Aya: From 101 Languages to 23
  6. North, Compass, and the Extended Family
  7. Pricing and Deployment
    1. API Pricing (per million tokens)
    2. Private Deployment
    3. Licensing Gotchas
  8. Competitive Position
  9. What This Means for Agent Builders
  10. Overall
  11. References

🌏 中文版

In 2019, Transformer co-author Aidan Gomez founded Cohere in Toronto. While most labs chased the strongest chat model, Cohere chose a different path — enterprise RAG infrastructure. Six years later that bet has grown into four pillars: Command for generation, Embed for retrieval, Rerank for precise ranking, and Aya for multilingual coverage. No competitor ships all four as product lines.

This is the eighth deep dive in the AI Model Family series, tracing Cohere from language-model startup to enterprise RAG suite. For how to read the benchmark numbers cited here, see our AI model evaluation sources guide. This is part of the AI Model Landscape Overview series.

Family Timeline

VersionReleaseKey fact
Cohere founded2019Toronto, by Aidan Gomez (Transformer co-author) with Nick Frosst and Ivan Zhang
Command (original)2022–2023Early generation models at $1.00/$2.00, establishing the API foundation
Command R2024-03-11First RAG-specialized model, 128K context, $0.50/$1.50
Command R+2024-04Flagship 104B params, 128K context, $3.00/$15.00, complex RAG and multi-step tool use
Command R+ 08-20242024-08Major refresh, +50% throughput, −25% latency, repriced to $2.50/$10.00
Aya 1012024-0213B, 101 languages, Apache 2.0
Aya 232024-058B/35B, 23 languages, CC BY-NC 4.0
Aya Expanse2024-12-058B (8K) / 32B (128K), 23 languages, CC BY-NC 4.0, API $0.50/$1.50
Command A2025-03-13111B, 256K context, runs on two GPUs, 150% higher throughput than R+ 08-2024
Embed v4.02025Mixed image-text (text + images + PDFs), 128K, variable dims 256–1536
Rerank v3.5 / v42025v3.5 single multilingual 4K; v4 Fast/Pro 32K with semi-structured JSON/YAML
North / Compass2025Enterprise workspace North + intelligent search Compass
North Mini Code2025First agentic coding MoE: 30B total / 3B active, 256K context, Apache 2.0
Transcribe2025Speech recognition, including Arabic fine-tuned variant

Six years, four generational leaps. Cohere's rhythm is not "a bigger model every quarter" but "each pillar evolves independently, converging on the North platform."

Four Pillars: Why Cohere Is Not Just an LLM Company

Most families have one main line (Qwen bets on full size coverage, DeepSeek on MoE efficiency, Claude on agents). Cohere runs four in parallel:

Command (generation): Chat, RAG grounded generation, multi-step tool use. From R to R+ to A, every generation doubled down on "how to do RAG with less compute" — Command A matches larger rivals' quality at 111B and needs only two A100/H100s.

Embed (retrieval): Turns text/images into vectors — the first stage of RAG. Before v4.0 it split into English vs. multilingual tracks (v3.0 full 1024-dim vs. light 384-dim); v4.0 unified into a single multilingual model with variable dimensions and multimodal input.

Rerank (precise ranking): The second RAG stage that re-scores retrieved candidates. v3.0 split English/multilingual, v3.5 unified to single multilingual, v4 extends to 32K and handles semi-structured data (tables, JSON, YAML). Billed per search unit (1 query + up to 100 docs).

Aya (multilingual): Research-driven multilingual series. Aya 101 covers 101 languages under commercially friendly Apache 2.0; Aya 23/Expanse narrows to 23 core languages with higher quality under CC BY-NC 4.0 (research-only).

The combination is Cohere's only moat: others make you assemble OpenAI + BGE + Jina + NLLB yourself; Cohere gives you the full stack from one vendor, one bill, one private deployment.

Architecture: Why "RAG-Native"

Grounded generation with citations

Command R/R+/A are trained from the start for "given documents, produce a cited answer." The model annotates where information came from rather than improvising. For enterprise knowledge-base QA — legal, support, finance — answers without sources are unacceptable.

Multi-step tool use

Since Command R+, the models support sequential multi-step tool use within a single response cycle, using prior tool outputs to decide the next step. This is the foundation for agents, and Cohere made it R+'s headline feature in 2024 rather than a late add-on. Command A adds emphasis on "avoiding unnecessary tool calls" — in practice, calling tools recklessly is more expensive than not calling them.

Efficiency-first design

Command A's pitch is not parameter count but efficiency: 111B params, 256K context, two GPUs, 150% higher throughput than R+ 08-2024. Cohere's official line is "on par or better than GPT-4o and DeepSeek-V3 across agentic enterprise tasks, with significantly greater efficiency" (Command A launch post). For enterprises that need private deployment, "the flagship runs on two cards" matters more than "2% higher on a benchmark."

Embed v4 and Rerank v4: How to Choose

Embed: From Dual Track to Unified

ItemEmbed v4.0Embed v3.0 (full)Embed v3.0 (light)
ModalityText + images + mixed PDFsTextText
Dimensions256 / 512 / 1024 / 1536 (default 1536)1024384
Context128K tokens512 tokens512 tokens
LanguagesSingle multilingual (100+)English / multilingual splitEnglish / multilingual split
DistanceCosine / Dot / EuclideanCosineCosine
Quantizationfloat / int8 / uint8 / binaryfloatfloat

Two key upgrades in v4.0: multimodal input (embed images and PDFs directly without converting to text first) and variable dimensions (pick 256 dims to save storage or 1536 dims to preserve quality). The old 512-token context was a hard ceiling for long documents; v4.0 at 128K can embed an entire document directly.

Pricing: no public per-token price for Embed — enterprise negotiation; Model Vault dedicated instances are Small $4/hr ($2,500/mo), Medium $5/hr ($3,250/mo).

Rerank: From 4K to 32K

ItemRerank v4 Pro / FastRerank v3.5Rerank v3.0
LanguagesSingle multilingualSingle multilingualEnglish / multilingual split
Context32K4K4K
Semi-structuredJSON/YAML/tables supported
BillingPer search unit (1 query + 100 docs)SameSame
Long docsAuto-chunked (>510 tokens)SameSame

The core differentiator in v4 is semi-structured data: enterprise documents mix tables, JSON, and YAML, and v4 Pro/Fast can score such content directly without converting to plain text. Bumping context from 4K to 32K allows reranking much longer documents in one pass.

Model Vault pricing: Rerank 3.5 and Rerank 4 Fast Medium both $5/hr ($3,250/mo), Rerank 4 Pro Large $10/hr ($6,500/mo).

One-Line Selector

  • New projects → v4: Embed v4.0 + Rerank v4 Pro is the current golden pair, covering multimodal and 32K reranking.
  • Cost-sensitive: Rerank 4 Fast and Pro cost the same at Medium ($5/hr) — pick Pro; for Embed use 512 dims instead of 1536 to halve storage and latency.
  • Already on v3: v3.0's English/multilingual split was unified in v3.5/v4 — upgrading simplifies integration.

Aya: From 101 Languages to 23

VersionParamsLanguagesContextLicenseAPI pricing
Aya 10113B (mT5-xxl)101Apache 2.0— (weight download)
Aya 238B / 35B238KCC BY-NC 4.0
Aya Expanse 8B8B238KCC BY-NC 4.0$0.50 / $1.50
Aya Expanse 32B32B23128KCC BY-NC 4.0$0.50 / $1.50

Aya's evolution shows convergence: from broad 101-language coverage to deeper quality on 23 core languages. Aya 101 remains the only commercially friendly Apache 2.0 release; Expanse offers higher quality and long context (32B reaches 128K) but under CC BY-NC for non-commercial use — enterprises needing commercial multilingual generation must go through the Cohere API rather than self-hosting weights.

The 23-language list: Arabic, Chinese (simp/trad), Czech, Dutch, English, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, Vietnamese — identical to Command A's 23.

North, Compass, and the Extended Family

Beyond models, Cohere layers two platforms:

North: Enterprise AI workspace bundling Command, Embed, Rerank, document connectors, and agent workflows. Think "ChatGPT + RAG + automation inside the enterprise," emphasizing private, secure integration with existing systems. Compass: Intelligent search and discovery, focused on document discovery and insight surfacing.

Extended family:

LineRepresentativePositioning
Flagship generationCommand A 111BEnterprise RAG, agents, multilingual, 256K, two GPUs
Previous flagshipCommand R+ 08-2024 104BStill active, 128K, $2.50/$10.00
Lightweight generationCommand R7BEdge / low-cost scenarios
RetrievalEmbed v4.0Multimodal, 128K, variable dims
RerankingRerank v4 Pro/Fast32K, semi-structured
Multilingual researchAya Expanse 32B23 languages, 128K
Open multilingualAya 101 13B101 languages, Apache 2.0
Agentic codingNorth Mini Code 30B-A3BMoE (30B total / 3B active), 256K, Apache 2.0
Speech recognitionTranscribe / Transcribe ArabicASR including Arabic fine-tune
Workspace platformNorth / CompassEnterprise workflows and search

Two observations:

North Mini Code marks Cohere's first step into coding. 30B total but only 3B active per token, 256K context, Apache 2.0 — similar MoE thinking to DeepSeek's approach, but Cohere's pitch is "a coding agent for the North platform" rather than a general coding leaderboard play.

Platformization cuts both ways. North packages the four pillars into a one-stop solution — a plus for enterprises already on Cohere; but for teams that only want Embed or Rerank, North's presence can raise "am I being locked in?" concerns. In practice, Embed/Rerank/Command are all usable standalone via API without going through North.

Pricing and Deployment

API Pricing (per million tokens)

ModelInputOutputNotes
Command R 03-2024$0.50$1.50Deprecated (2025-09-15)
Command R+ 04-2024$3.00$15.00
Command R+ 08-2024$2.50$10.00Current workhorse
Command A 03-2025$2.50$10.00Same price as R+ 08-2024, more efficient
Aya Expanse 8B/32B$0.50$1.50
Embed v4.0Custom enterpriseNo public per-token price
RerankPer search unit1 query + 100 docs per unit, >500 tokens auto-chunked

Sources: Cohere Pricing and Command A docs.

Private Deployment

Two private paths, both Kubernetes-containerized, available on AWS/Azure/GCP/OCI:

  • Private Deployments: Runs in the customer's own environment (on-prem or isolated VPC), data never leaves.
  • Model Vault: Cohere-managed dedicated instances (no multi-tenancy), in Standard and Encrypted (confidential computing, remote attestation) tiers, compliant with GDPR, HIPAA, and SOC 2.

Model Vault is billed per instance-hour; see the Embed/Rerank tables above for those Vault prices. Vault pricing for the Command series requires contacting sales.

Licensing Gotchas

Cohere's licensing splits into three tiers — check before you choose:

  • API (Command/Embed/Rerank): Commercial API, pay-as-you-go, no weight downloads. Data-sovereignty-sensitive workloads require Private Deployment, not "download and self-host."
  • Apache 2.0 (Aya 101, North Mini Code): Commercially usable, self-hostable, modifiable. The only two truly open releases from Cohere.
  • CC BY-NC 4.0 (Aya 23, Aya Expanse, Command R+/A HF weights): Research and non-commercial only; commercial use must go through the API. All CohereLabs/c4ai-command-* and aya-expanse-* weights on HuggingFace fall here.

Compared to other families in this series: Qwen (Apache 2.0/custom), DeepSeek (MIT), Llama 4 (Community License), and Mistral (Apache/Modified MIT) all offer commercially self-hostable weights; Cohere's flagship generation and retrieval models do not — going private means going through official deployment channels.

Competitive Position

Cohere in the 2026 landscape:

  • vs. OpenAI (GPT-5.6 + text-embedding-3): GPT leads on general quality and ecosystem, but Cohere's Embed/Rerank are purpose-built for RAG with grounded generation and citations. For "traceable enterprise QA," Cohere is the lower-friction choice.
  • vs. BGE-M3 / Jina (open retrieval): BGE-M3 (MIT, 36M downloads) + Jina Reranker v3.5 is the default open RAG stack — self-hostable and cheap. Cohere Embed v4/Rerank v4 wins on multimodal, 32K, semi-structured, and enterprise support, not raw scores.
  • vs. Qwen / DeepSeek (open generation): Both offer commercially self-hostable weights (Apache 2.0/MIT); Cohere Command does not — but Cohere's multi-step tool use and grounded RAG are more mature for enterprise workflows.
  • vs. Claude / Gemini (closed generation): Claude Opus 5 and Gemini 3 lead on coding and reasoning; Cohere's differentiation is the "generation + retrieval + reranking" bundle with private deployment, not single-point quality.
  • vs. NLLB / SeamlessM4T (multilingual/translation): Aya 101's 101-language coverage matches NLLB-200 and adds instruction following (context-aware translation); Aya Expanse delivers stronger generation within 23 languages.

One-line positioning: If you want a single /chat/completions endpoint, Cohere is not the first pick; if you want a privately deployable RAG pipeline, Cohere is one of the few families that ships it complete.

What This Means for Agent Builders

  • End-to-end RAG pipeline → Embed v4 (recall) + Rerank v4 Pro (rerank) + Command A (cited generation) is Cohere's golden path — one API vendor, one private deployment.
  • Multi-step agents → Command A/R+ multi-step tool use supports sequential calls ("search → read docs → compute → respond").
  • Multilingual agents → Aya Expanse 32B (128K) handles cross-lingual documents, or Aya 101 (Apache 2.0) for self-hosted multilingual classification/translation.
  • High-throughput / cost-sensitive → Embed and Rerank APIs are optimized for retrieval latency and cost vs. "using an LLM as a reranker"; Aya Expanse at $0.50/$1.50 is also in the low-price tier.
  • Private / compliant → Model Vault (Standard/Encrypted) and Private Deployments keep data on-prem — a top pick for finance, government, and healthcare — but flagship generation has no commercially self-hostable weights, so vendor lock-in is part of the evaluation.
  • Local / edge → Only Aya 101 and North Mini Code are Apache 2.0 self-hostable; for local Embed/Rerank consider open alternatives like BGE-M3 + Jina.

Pragmatic mix-and-match:

TaskRecommendationWhy
Enterprise knowledge-base QACohere Embed v4 + Rerank v4 Pro + Command ANative grounded generation with citations
Multilingual doc processingAya Expanse 32B (API) or Aya 101 (self-hosted)101/23 language coverage, instruction-driven translation
Low-cost multilingual generationAya Expanse 8B API $0.50/$1.50Cheapest multilingual generation at this quality
Local/offline RAGBGE-M3 + Jina Reranker + Qwen/DeepSeekCohere retrieval has no commercially self-hostable weights
Complex coding agentClaude Opus 5 / DeepSeek V4 Pro + Cohere RerankStrongest generator + Cohere reranking
Agent needing private deploymentCohere Private Deployment / Model VaultData stays on-prem, SOC 2/HIPAA/GDPR

Overall

Cohere proves a different survival strategy: instead of competing with OpenAI and Anthropic on "the strongest chat model," it productizes every stage of RAG. Where rivals offer an endpoint, Cohere offers a pipeline — Embed for recall, Rerank for precision, Command for generation, Aya for crossing languages, all runnable in a private environment.

The cost is openness. Flagship generation and retrieval have no commercially open weights, and Aya's openness stepped back from Apache 2.0 to CC BY-NC. Cohere's "open" is research-open, not deployment-open. Before choosing Cohere, confirm you need "a RAG infrastructure you can buy" rather than "weights you can self-host" — that line matters more than any benchmark score.


References