What Is Cohere Command A+ and How Does It Work?
Cohere Command A+ is a large language model from Cohere with 218 billion parameters. It is designed to run on only two H100 GPUs, making it a powerful but efficient option for businesses. Released on May 20th, it uses a mixture-of-experts (MoE) architecture that activates only 25 billion parameters at a time.
Think of it like a hospital with many specialists. When you have a broken arm, you don't see every doctor. You see the right one. Command A+ does the same: only the relevant parts of the model turn on for your request. This keeps it fast and affordable while still having access to its full knowledge.
Key Features of Command A+ You Should Know
Several features make Command A+ stand out. First, it is fully open source under the Apache 2.0 license. That means you can download it, run it, and use it commercially without restrictions. Second, it is multimodal—it can process images and documents, not just text.
- 218 billion parameters total, but only 25 billion active at a time
- Runs on two H100 GPUs instead of a whole data center
- W4A4 quantization for lossless compression
- Native citations to reduce hallucinations
- Supports 48 languages, including all EU official languages
- Ranked first on the AARF Omni non-hallucination benchmark with 86%
These features combine to give you frontier-level AI without massive hardware costs. Cohere co-founder Nick Frost said the tech can empower people, and that is the goal.
How to Get Started with Command A+
Getting started with Command A+ is simple because it is open source. You can download the model today from Hugging Face. No special permissions are needed.
- Go to Hugging Face and search for Command A+.
- Download the model weights to your own servers or machines.
- Make sure you have at least two H100 GPUs available to run it efficiently.
- Set up your environment with the necessary libraries, like PyTorch or Hugging Face Transformers.
- Load the model and start making API calls locally.
You can also have a team deploy it for you if you prefer. This gives you full control over your data and avoids paying per-token API fees.
Using Command A+ in Real Business Examples
In the episode, a real example shows how Command A+ can research and create a comparison document. Suppose you want to find the best AI automation communities for business owners. You feed the model articles, forum threads, and reviews.
With native citations, it gives you an answer and shows which source backs each claim. So if you were building a document to compare options, every fact would be sourced. This turns a task that used to take hours into a minutes-long process.
- Research with verified citations
- Create comparison reports quickly
- Process images and documents for data extraction
- Translate content into 48 languages
This makes Command A+ useful for content creation, data analysis, and customer support.
Performance and Benchmark Results Explained
Command A+ does not beat every model on every test, but it competes well on math and reasoning. It trails models like DeepSeek V4 Pro on some coding and general intelligence benchmarks. However, the focus is on efficiency.
On the AARF Omni non-hallucination benchmark, Command A+ scored 86%, about 3 percentage points higher than the next best model. That is important for businesses because fewer hallucinations mean more reliable outputs.
Cohere calls its quantization lossless—performance barely drops after compression. That is why the model can run on two GPUs instead of eight or sixteen. The trade-off is worth it for most business applications.
Tips, Limits, and Mistakes to Avoid
When using Command A+, remember it is not the best at everything. It still loses to bigger models on some tasks. Do not expect it to outperform every frontier model on coding or general intelligence.
Avoid relying on it for tasks where you need the absolute highest accuracy on complex coding, unless you test it first. The model is efficient, but you still need proper GPU setup.
- Ensure you have two H100 GPUs; fewer may cause performance issues.
- Use citations to verify facts, but still double-check important claims.
- Start with simple tasks to understand its behavior.
- Monitor hallucinations even with the low rate.
Also, remember that it is open source, so you are responsible for maintaining it and updating it.
Who Should Use Command A+ and Why It Matters
Command A+ is built for enterprises, governments, and institutions that want sovereign AI—keeping data on their own infrastructure. Fujitsu already said its architecture fits their sovereign AI strategy. This is a big shift from relying on third-party APIs.
For business owners, this means you can have frontier-grade AI without paying huge cloud costs. You are not locked into one provider. You can deploy it yourself and build workflows that save hours weekly.
The model is accessible now, and the AI space is moving fast. Business owners who learn to use tools like this early will benefit most.
