Cohere Command A+

Command A+ is Cohere's first mixture-of-experts model and the current top of the Command family, released 2026-05-20. It pairs a strong agentic and reasoning profile with vision input and a jump to 48 languages, which is the widest language coverage Cohere has shipped in a Command model and includes every official European Union language.

The architectural point is efficiency per instance. With 218B total parameters and 25B active, Command A+ is designed so an enterprise can serve it on one B200 or two H100s at W4A4 rather than a rack. That is the whole argument for MoE at this scale: you keep the accuracy of a large model and pay roughly the compute of a much smaller one per token.

Model Specifications

Specification	Command A+
Architecture	Sparse mixture of experts
Parameters	218B total, 25B active
Context window	192K tokens
Input	Text, image
Output	Text
Languages	48, including all official EU languages
Required hardware	1x B200 or 2x H100 at W4A4
License	Apache 2.0

Unlike most vendors at this tier, Cohere publishes the weights: the checkpoint is on Hugging Face as CohereLabs/command-a-plus-05-2026-w4a4 under Apache 2.0, which allows self-hosting and fine-tuning without a commercial licence negotiation.

Where it fits

Cohere's own framing is enterprise agentic work: complex multimodal tasks where the model takes actions in an environment, and multilingual business workloads where 40-plus languages are a requirement rather than a nice-to-have. The model supports native tool calling with strict tool schemas, structured outputs, and a reasoning mode with token budgets, so it can be dropped into pipelines that need guaranteed JSON shapes rather than free text.

If you are already on Command A, Command A+ is the upgrade path rather than a parallel product. Cohere also ships a private deployment story: Managed Vault, encrypted Vault, on-prem or isolated VPC, which is the part that usually decides Cohere against a pure API vendor.

Pricing

Cohere sells through its own platform and through partners, so list prices vary by deployment type. The managed listing on OpenRouter, checked on 2026-09-23, was:

Item	Price
Input	$0.30 / 1M tokens
Output	$1.50 / 1M tokens
Cached input	$0.15 / 1M tokens

Self-hosted costs instead land on your GPU bill, where a single B200 or a pair of H100s is the floor Cohere targets.

Limitations

192K context is mid-pack in 2026. The current frontier tier runs 500K to 1M tokens; see Qwen3.8-Omni-Flash for the long-context end of this class.
Text and image input, text output only. No audio or video, and no image generation.
No independent index score posted yet. OpenRouter carried Artificial Analysis scores for the older Command A but not for Command A+, so treat capability claims as vendor positioning for now.
Enterprise-shaped, not consumer-shaped. There is no free chat tier that tells you much about daily use.

FAQ

Is Command A+ open source?

The weights are open under Apache 2.0 on Hugging Face, and they are also available through Cohere's own deployments and partners. That combination is unusual: you can self-host or buy it managed.

How is it different from Command A?

Command A is the previous generation and the first MoE design was reserved for A+. A+ adds vision input, widens language support to 48, and improves reasoning and agentic behaviour over the rest of the family.

Does Cohere have embedding and rerank models too?

Yes. Cohere's retrieval stack includes Cohere Embed v3 and Cohere Rerank 3.5, which is often why teams adopt its generative models in the first place.

Alternatives

Command A: the previous generation, still listed for existing integrations.
Qwen3.8-Omni-Flash: far longer context and audio-video input, but closed weights.
MiMo V2.6: open-weight multimodal alternative when you want an MIT-licensed option.