What is Command A+?
Command A+ is the newest model in Cohere's Command A family and its first mixture of experts, with 218B parameters in total and 25B active per token. It takes text and images, writes text, reasons before answering and calls tools, across a 128K-token input with up to 64K tokens out; its knowledge stops at April 2025.
Cohere built it from a year of running North, its agent workspace, and folds the separate Reasoning, Vision and Translate models into one set of weights. Unlike Command A, whose weights were for non-commercial use, it ships under Apache 2.0, so companies can run it on their own GPUs.
What changed from Command A?
It trades Command A's longer context for image input, twice the languages, a mixture-of-experts design and a licence that allows commercial use.
- Licence
- Apache 2.0, where Command A's weights were for non-commercial use only.
- Design
- A mixture of experts with 218B parameters and 25B active per token, against Command A's 111B dense model.
- Languages
- 48, up from 23.
- Images
- Image input is built in; before, that needed the separate Command A Vision model.
- Context
- 128K tokens of input, down from Command A's 256K.
- Speed
- Up to 63% more output tokens per second than Command A Reasoning, by Cohere's measure.
Who is Command A+ best for?
Command A+ is best for companies that self-host their agents, work across many languages, read scanned documents and charts, or want to test a Cohere model for free before buying GPUs.
Self-hosted enterprise agents
The 4-bit weights run on two H100s or a single B200, so a company can keep its agent model on its own servers.
Work across many languages
It covers 48 languages, including every official EU language, against 23 in the earlier Command A models.
Reading documents and charts
It takes images alongside text, so one model can read scanned pages and charts, reason over them and then call tools.
Trying it before buying hardware
A free Cohere API key gives 20 requests a minute and 1,000 calls a month to test it before you commit GPUs or a Model Vault contract.
What's included with Command A+?
Switchable reasoning, tool use with citations, image input, 48 languages, three downloadable weight formats, and dedicated hosting on Model Vault.
- Reasoning you can switch off
- It thinks before it answers by default, and the thinking step can be turned off for quicker replies.
- Tool use with citations
- Built for agents that call APIs, databases and search, with cited answers, structured JSON output and safety modes.
- Image input
- Reads charts, scanned pages and photos alongside text; Cohere reports 80.6% on MathVista, up from 73.5% for Command A Vision.
- 48 languages
- Covers every official EU language plus Arabic, Japanese, Korean, Hindi and more, with a new tokenizer that needs fewer tokens for Arabic, Japanese and Korean.
- Three weight formats
- BF16, FP8 and 4-bit downloads that Cohere says score almost the same; the 4-bit version is the one it recommends.
- Model Vault hosting
- Cohere can run it for you on a dedicated, isolated instance billed by the hour.
How much does Command A+ cost?
Command A+ is free to download under Apache 2.0 and free on a Cohere API key up to 20 requests a minute and 1,000 calls a month; hosted production use runs on Cohere's Model Vault from $17.50 per instance per hour.
- 20 requests a minute
- 1,000 API calls a month
- Not for production or commercial use
- BF16, FP8 and 4-bit weights on Hugging Face
- 4-bit runs on two H100s or one B200
- Commercial use allowed
- Dedicated instance that Cohere manages
- Access usually through a waitlist
- Monthly and yearly commitments priced by sales
- Larger performance tier
- Fixed or autoscaling Flex plans
Cohere's pricing page lists Command A+ at $0 for an API key and for the download; the Model Vault hourly rates come from Cohere's Standard Vault pricing, and monthly or yearly rates come from its sales team. Production API keys also run Command A+ free up to the rate limits, but Cohere sends production use to Model Vault. It is also on Microsoft Azure AI Foundry at Azure's own prices. Prices are in US dollars. Full pricing on cohere.com. We checked these prices on .
How do readers rate Command A+?
No one has rated Command A+ yet. Be the first.
Sign in to rate. One rating per account.
Command A+ pros and cons
Its Apache 2.0 weights fit on two H100s and beat earlier Command models on Cohere's tests, but the context is only 128K and production hosting is not pay-per-token.
Pros
- Its Apache 2.0 licence allows commercial self-hosting, which fixes the main complaint Reddit users had about earlier Command models and their non-commercial licences.
- Fits a 218B-parameter model on two H100s or one B200 with 4-bit weights that Cohere says lose almost no quality.
- Large gains over Command A Reasoning on Cohere's tests: 85% against 37% on tau2-Bench Telecom, 25% against 3% on Terminal-Bench Hard, and up to 63% faster output.
- One set of weights covers reasoning, tool use, image input and 48 languages, work that needed separate Command A models before.
Cons
- Its 128K-token input is half of Command A's 256K and short next to open rivals such as Mistral Large 3 (256K) or Nemotron 3 Super (1M), and its knowledge stops at April 2025.
- There is no simple pay-per-token production API: trial keys cannot be used commercially, and hosted production use means a Model Vault instance from $17.50 an hour, usually through a waitlist, or your own GPUs.
- Local users on Reddit found the 4-bit weights need about 132 GB of memory and do not yet load in Apple's MLX, so it still needs data-centre GPUs or a very large Mac.
Is Command A+ open source?
Yes. Command A+'s code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Command A+ alternatives?
Command A+ is an open source alternative to Claude Sonnet 5 and GPT-5.5. DeepSeek-V3.2-Speciale, GLM-4.7, DeepSeek V4 Pro and LongCat-2.0 do the same job and are open source too.
Claude Sonnet 5FreeAnthropic's Sonnet 5 with a 1M context, at $2 and $10 per million tokens.
GPT-5.5PaidOpenAI's agentic model for coding and professional work, with a Pro tier.
DeepSeek-V3.2-SpecialeOpen SourceDeepSeek's high-compute V3.2 variant for hard math and coding problems, released as open weights.
GLM-4.7Open SourceZ.ai's MIT-licensed model for agentic coding with a 200K context.
DeepSeek V4 ProOpen SourceDeepSeek's 1.6T-parameter open MoE model for coding and agents, with a 1M-token context.
LongCat-2.0Open SourceMeituan's open-weight 1.6T-parameter MoE model with a 1M-token context for coding and agent work.
Questions about Command A+
What is Command A+?
Command A+ is Cohere's open-weight model for enterprise agents: it reads text and images in 48 languages, reasons and calls tools, and is free to download under Apache 2.0.
Who is Command A+ for?
Command A+ is best for companies that self-host their agents, work across many languages, read scanned documents and charts, or want to test a Cohere model for free before buying GPUs.
Is Command A+ free?
Yes. Command A+ has a free API key plan: 20 requests a minute.
How much does Command A+ cost?
Command A+'s paid plans are Model Vault L at $17.50 per instance per hour and Model Vault XL at $32.50 per instance per hour.
Is Command A+ open source?
Yes. Command A+'s code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Command A+ alternatives?
The closest alternatives on Recaply are Claude Sonnet 5, GPT-5.5, DeepSeek-V3.2-Speciale, GLM-4.7, DeepSeek V4 Pro and LongCat-2.0.
Where can I try Command A+?
On its official website, cohere.com.
Can I use Command A+ in a commercial product?
Yes, when you host it yourself, since Cohere releases the Command A+ weights under the Apache 2.0 licence. Cohere's free trial API keys are not allowed for production or commercial use, so hosted production use goes through Cohere's Model Vault, which starts at $17.50 per instance per hour.
What hardware does Command A+ need?
Command A+ runs on as little as two NVIDIA H100 GPUs or one B200 with the 4-bit weights, which Cohere recommends for most uses. The FP8 version needs four H100s or two B200s, and the full BF16 version eight H100s or four B200s.
Is Command A still available now that Command A+ is out?
Yes. Cohere still lists Command A as live on its API, with a 256K context and 500 requests a minute on production keys. Command A+ is the newer model, with image input, 48 languages and Apache 2.0 weights, but half the context.
Is Command A+ on AWS or Azure?
Cohere lists Command A+ on Microsoft Azure AI Foundry as a 4-bit deployment, but not on Amazon Bedrock, Amazon SageMaker or Oracle's generative AI service. Elsewhere you can call Cohere's own API, rent a Model Vault instance, or run the Hugging Face weights on your own cloud GPUs.
Does Cohere use my Command A+ data for training?
Not when you host it yourself, since Cohere has no access to inputs in your own environment. On Cohere-hosted services, Cohere says de-identified usage data may be used in limited cases where user controls and its terms allow, so check those settings before sending sensitive data.
Our verdict
Command A+ is a good pick for companies that must run an agent model on their own GPUs and need many languages, image input and a licence that allows commercial use.
It fits on two H100s at 4-bit, Cohere reports large gains over earlier Command models, and the Apache 2.0 weights remove the non-commercial limit that held Command A back.
- Skip it for
- Documents longer than 128K tokens, and teams that just want a cheap pay-per-token API, since production hosting means Model Vault or your own servers.
- Check first
- Test the 4-bit weights on your own serving stack first; Reddit users found they need about 132 GB of memory and do not yet load in Apple's MLX.
Muse Spark
Kimi K2.5
Ring-2.6-1T
Gemma 4
Kimi K3
LFM3