Cohere Command A+ Review: 218B Sparse MoE, Apache 2.0, and Native Citations Built In

At a glance: Cohere Command A+, released May 20, 2026. 218B sparse MoE, 25B active parameters per token. Apache 2.0 license. Native citation generation. No public per-token API pricing published by Cohere. Artificial Analysis Intelligence Index: 37 at launch, 23 as of this audit under AA’s revised methodology. Knowledge cutoff April 1, 2025. Part of our AI Models & Companies reviews.

Cohere’s most important announcement in Command A+ has nothing to do with a benchmark.

Cohere’s prior model, Command R+, shipped under CC-BY-NC 4.0 — meaning commercial deployment required an enterprise license from Cohere. You could evaluate it, research with it, and integrate it in non-commercial contexts, but the moment you shipped a product, you needed a licensing conversation. Command A+ drops that requirement entirely. Apache 2.0. Full commercial use. No revenue caps. No non-compete clauses. Take the weights, fine-tune on your classified data, deploy on an air-gapped server, ship a product. Cohere gets nothing from you except attribution in the license file.

That licensing shift, paired with a 218B sparse MoE architecture that runs on two H100 GPUs, makes Command A+ the most practically deployable open-weight enterprise model in the market. Cohere has not published public per-token API pricing for Command A+ (see Pricing below); whether the self-hosting economics or a custom enterprise deal make sense — against alternatives that are cheaper or more capable on raw benchmarks — depends on what you’re building.

Release Context

Command A+ was announced May 20, 2026 in a blog post titled “Introducing Command A+: Our Most Open, Capable, and Efficient Model Yet.”

The release is a substantial departure from Cohere’s prior release cadence. Command A (the predecessor) was released in March 2025 under CC-BY-NC 4.0 — the same terms as Command R+. Command A Reasoning, a thinking-mode variant, arrived later under similar restrictions. Command A+ collapses both into a single model with a fundamentally different license.

Predecessor comparison:

The licensing transition is the headline. The multimodal addition — Cohere’s first — is the capability headline. Both matter.

Architecture

Sparse MoE with Selective Quantization

Command A+ uses a sparse Mixture-of-Experts decoder with 218 billion total parameters organized across 128 experts. Each token activates 8 experts plus 1 shared expert — meaning 25 billion parameters are active per inference call, not 218 billion. This is the same architectural efficiency principle as Mistral Small 4 (6B of 119B active) or DeepSeek V4 (49B of 1.6T active): total parameter count overstates the compute cost of any single forward pass.

The genuinely innovative piece in Command A+ is Quantization-Aware Distillation (QAD): rather than applying uniform quantization across the model, QAD preserves attention pathway weights at full precision while quantizing only the MoE expert layers. The result is near-lossless W4A4 quantization — Cohere claims the compressed model retains quality while running on significantly less hardware. Independent verification of “near-lossless” claims will take time to accumulate from the research community, but the hardware footprint reduction is concrete.

W4A4 performance numbers (Cohere-reported, independently confirmed by VentureBeat):

• Output throughput: 375 tokens/second at low concurrency
• Time-to-First-Token: 113ms
• Hardware: 2×H100 or 1×Blackwell B200

For comparison: Command A Reasoning ran at ~237 tokens/second on the same inference setup. The QAD-compressed Command A+ is 63% faster than its predecessor while fitting on fewer GPUs.

Specifications

The Apache 2.0 Licensing Story

The license is what builders should read carefully before anything else.

Apache 2.0 means:

• Commercial use with no restrictions
• Modification and redistribution permitted
• No revenue caps (unlike Mistral’s $20M/month Modified MIT)
• No non-compete clauses (unlike some Meta Llama licenses)
• Fine-tuning and derivative models permitted under same terms
• Air-gapped deployment — weights, quantization files, and all artifacts on HuggingFace are fully downloadable

What this unlocks for enterprise:

The practical unlock is in regulated industries. A hospital system building clinical decision support can fine-tune Command A+ on proprietary patient data without those data ever leaving their infrastructure. A financial institution can deploy Command A+ on hardware that has never touched the public internet. A defense contractor can run Command A+ on a classified network. None of these use cases were legally clean with CC-BY-NC 4.0 without a separate enterprise agreement.

Comparison to the competition:

Command A+ and Mistral Large 3 share the most open licensing profile among frontier-adjacent models. What differentiates them is Cohere’s enterprise feature set: primarily the native citation system.

Native Citations

This is the genuinely unique feature in Command A+, and it’s worth explaining in detail.

Most RAG pipelines handle citation attribution as a post-processing step: the model generates an answer, then a separate component matches phrases in the answer to source chunks. This is fragile — the phrase matching can fail, the model can generate plausible-sounding content that doesn’t map cleanly to a source, and the citation layer adds latency and infrastructure complexity.

Command A+ generates citations during inference using <co> / </co> tags that wrap factual claims directly in the output. Each tag links to a source document or database row by reference. There is no post-processing step — the model is trained to attribute as it generates.

Practical implications:

• Every factual claim in a response carries a machine-readable citation to its source
• Suitable for direct use in legal documents, clinical notes, financial reports — contexts where unsourced AI-generated claims are a liability
• Reduces the engineering complexity of building citation-aware RAG pipelines
• Cohere reports 20% improvement in “Agentic Question Answering accuracy” over Command A Reasoning, partly attributed to multimodal document understanding feeding into more accurate citations

For teams building in regulated industries, this is a meaningfully different capability from what other open-weight models provide. No other major open-weight model ships with architecture-level citation generation trained into the model itself.

Benchmarks

Agentic Performance: The Highlight Numbers

These are generation-over-generation improvements versus Command A Reasoning, reported by Cohere, not absolute comparisons to frontier models. The τ²-Bench Telecom jump from 37% to 85% is the headline: Cohere’s own agentic benchmark shows a substantial capability increase for sustained multi-step tool use.

Broader Benchmarks

What the Intelligence Index Tells You

At launch, Command A+ scored 37 on the Artificial Analysis Intelligence Index (May 20, 2026) — in line with Claude 4.5 Haiku and just above NVIDIA Nemotron 3 Super at the time. Update (August 2026): Artificial Analysis has since revised its Intelligence Index methodology (v4.1, introduced June 2026), which recalculated scores across every model on the index — not something specific to Command A+. As of this audit, Command A+ scores 23 on the current version of the index. Current scores for the same comparison set:

The relative story hasn’t changed even though the absolute numbers have: Command A+ still trails Mistral Medium 3.5 on the composite index, and remains well below the frontier cluster. For tasks that require deep reasoning or graduate-level science (11% on HLE), Command A+ is not the best available option. Cohere’s own launch materials do not include a dedicated coding benchmark for Command A+, so builders evaluating code-heavy use cases have no first-party signal here and should benchmark independently rather than assume parity with frontier coding models.

The benchmark profile is consistent with a model optimized for enterprise tool use and RAG workflows rather than raw reasoning capability. That’s a legitimate product positioning — just not a frontier-tier general-purpose claim.

Pricing

Cohere API: As of this audit (August 2026), Cohere has not published public per-token pricing for Command A+. It does not appear on Cohere’s public pricing page (which lists only legacy models like Command and Command R+ 08-2024 at public rates), and Cohere’s own model documentation does not list a price for it either — access to hosted inference is arranged directly with Cohere rather than through a published self-serve rate card. Some third-party cost trackers list $2.50 input / $10.00 output per million tokens for Command A+, but that figure is identical to Command R+ 08-2024’s published rate and could not be independently confirmed as Command A+'s own pricing from a Cohere source; treat it with caution until Cohere publishes a rate card for this model.

(Table corrected 2026-08-23: the prior version of this table understated Command R+'s current rate, overstated Claude Opus 4.7’s rate roughly 5x, understated GPT-5.5’s output rate, and stated a Command A+ price that Cohere’s own pricing and docs pages do not support. See sources linked in each row.)

Without confirmed API pricing for Command A+ itself, a direct cost comparison against these alternatives isn’t possible from public information. What is clear: self-hosting the Apache 2.0 weights sidesteps the question entirely.

For teams self-hosting: inference cost is whatever your hardware costs. At W4A4 on 2×H100, the marginal cost of additional output tokens is compute amortization, not a metered API fee. Self-hosting is where the Apache 2.0 license and efficient quantization combination delivers the most value — and, per Cohere’s own materials, appears to be the primary way Command A+ is actually being accessed.

Self-Hosting

The “runs on 2×H100” headline is accurate for the W4A4 quantized version. Practical details:

• BF16 full precision: ~8×H100 80GB (full precision MoE inference requires more memory due to all expert weights being loaded)
• FP8: ~4×H100 80GB
• W4A4 (QAD quantized): 2×H100 80GB or 1×NVIDIA B200 — the deployment target for the headline claim

At W4A4, Cohere reports 375 tokens/second output throughput and 113ms time-to-first-token. For most agentic applications (tool use, RAG, document processing), 375 t/s is more than adequate.

Available on HuggingFace (CohereLabs):

• BF16 weights: CohereLabs/command-a-plus-05-2026
• FP8: CohereLabs/command-a-plus-05-2026-fp8
• W4A4: quantized via QAD (Cohere’s tooling)

vLLM and standard Transformers inference are supported from day one. Azure AI Foundry integration is listed at launch (deployment ID coherelabs-command-a-plus-05-2026-w4a4); Cohere’s own dashboard, AWS Bedrock, SageMaker, and Oracle listings show it as not yet available or coming soon.

Multimodal

Command A+ is Cohere’s first model to accept image input. The vision capability is focused on document and spreadsheet processing — the 32% improvement in spreadsheet analysis and 20% improvement in agentic QA accuracy are the multimodal capability claims Cohere leads with.

This is not a general image understanding model in the style of GPT-4o or Gemini 3.1 Pro. The MMMU scores (75.1%) are competitive but not leading. The practical use case is: upload a PDF, a chart, or a spreadsheet alongside a text query, and Command A+ will reason across both. For enterprise document workflows, that’s the capability that matters. For general multimodal tasks, frontier alternatives are more capable.

Comparison to the Cohere Ecosystem

vs. Command R+ (the predecessor): Command A+ is faster, multimodal, more capable on agentic benchmarks, more permissively licensed, and supports 48 vs. 23 languages. Command R+ should be retired for new projects.

vs. Command A Reasoning: Command A+ replaces Command A Reasoning as the flagship. The capability improvements (τ²-Bench +48 points, AIME +33 points) are substantial. The license upgrade (Apache 2.0 vs. CC-BY-NC) is the procurement argument.

The self-hosting thesis: Cohere’s bet with Command A+ is that enterprise builders who need data sovereignty and citation-quality RAG will self-host rather than use the API. Apache 2.0 + W4A4 efficiency + native citations + 48 languages is a package designed for that buyer. The API exists for teams that want managed inference without the hardware commitment.

Who Should Use Command A+

Strong fit:

• Regulated industry RAG pipelines (legal, finance, healthcare) that need machine-readable citations on factual claims
• Sovereign AI deployments requiring air-gapped or on-premise infrastructure with no vendor agreement
• Multilingual enterprise applications (48 languages, including several not well-served by US-headquartered models)
• Teams whose primary bottleneck is licensing flexibility, not raw reasoning capability

Weak fit:

• Coding-heavy applications — Cohere has not published a coding-specific benchmark for Command A+; teams should test it directly against GPT-5.5 or Claude Opus 4.7 rather than assume parity
• Graduate-level scientific reasoning — HLE at 11% is a significant gap vs. frontier models
• Latency-sensitive consumer applications — 113ms TTFT is competitive but not leading
• Knowledge cutoff sensitive applications — April 2025 is 13 months old at launch; if recency matters, check alternatives

Rating: 3.5/5

What earns the 3.5:

The Apache 2.0 licensing is a genuine unlock — particularly for regulated industries and government deployments where CC-BY-NC required a separate agreement. Native citations are architecturally unique in the open-weight space: no other comparable model ships citation generation as a trained capability rather than a post-processing layer. The QAD quantization achieving 375 t/s on 2×H100 is a real infrastructure story, not a marketing claim. τ²-Bench Telecom at 85% (from 37%) represents a substantial agentic capability improvement for multi-step tool use.

What limits the score:

The AI Intelligence Index places Command A+ in the capable-but-not-frontier tier — 37 at launch (in the same band as Mistral Medium 3.5’s then-39), and 23 as of this audit under Artificial Analysis’s revised v4.1 methodology (Mistral Medium 3.5 now 30; the frontier cluster of GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro now sits around 48-55). Either way, Command A+ trails Mistral Medium 3.5 and sits well below frontier models on this composite. The knowledge cutoff of April 1, 2025 was already 13 months old at launch — a meaningful gap for applications requiring current information. Cohere has not published public API pricing for Command A+, so a head-to-head cost comparison with Mistral Medium 3.5 ($1.50/$7.50) or other hosted alternatives isn’t possible from public information. And Cohere has not published coding benchmarks for Command A+ — the most common reason builders reach for open-weight models — leaving that gap unaddressed rather than explicitly conceded.

Command A+ is the right model for a specific buyer profile: enterprise teams who need data sovereignty, machine-readable citations, and Apache 2.0 licensing for regulated industry deployments. For everyone else, the capability-to-cost profile relative to DeepSeek V4, Mistral alternatives, or the Qwen 3.x family requires more justification.

Quick Reference

Sources

• Introducing Command A+: Our Most Open, Capable, and Efficient Model Yet — Cohere blog, May 20, 2026 — architecture, benchmarks, licensing, hardware requirements
• CohereLabs/command-a-plus-05-2026 — HuggingFace — model weights, architecture details
• Command A+ — docs.cohere.com — model card: context length, hardware, Azure AI Foundry availability, no published pricing
• Cohere pricing page — confirms no public Command A+ rate; legacy Command R+ 08-2024 rate ($2.50/$10.00 per M tokens)
• Cohere cracks lossless quantization and native citations… — VentureBeat, May 20, 2026 — independent confirmation of 375 tok/s throughput, 113ms TTFT, AIME 25, HLE, GPQA Diamond scores
• Cohere launches open weights model Command A+ — Artificial Analysis, May 20, 2026 — launch-day Intelligence Index score (37)
• Command A+ — Artificial Analysis Intelligence Index (current) — current composite score (23) under AA’s revised methodology
• Artificial Analysis Intelligence Index v4.1.1 methodology — explains the June 2026 index-wide recalculation
• Current Intelligence Index scores for comparison models: Claude Opus 4.7, GPT-5.5 (high), Gemini 3.1 Pro Preview, Mistral Medium 3.5
• Current competitor API pricing: OpenAI GPT-5.5, Anthropic Claude Opus 4.7, Mistral Medium 3.5, DeepSeek V4 Pro

Related ChatForest reviews: Cohere Enterprise Platform · Mistral Medium 3.5 · Mistral Large 3 · DeepSeek V4 · Mistral Small 4