Command A Guide: Cohere's 111B Dense Model and Its Licence Catch

Command A is easy to misfile. The weights are on Hugging Face, so people file it as an open model they can build a product on. They cannot, at least not without talking to Cohere first, and that is the single most important fact about it for anyone planning a deployment.

The second most important fact arrived in May 2026, when Cohere released Command A+, which it says unifies the capabilities of the earlier Command A models. The question for Command A is now less "is it good" and more "is there still a reason to pick it over its successor". There is, but it is narrower than it was.

What Command A is

Cohere released Command A in March 2025 under the API name command-a-03-2025. It is a dense 111B-parameter model with a 256K-token context window, trained with supervised fine-tuning and preference training. The model card lists 23 languages, including English, French, Spanish, German, Chinese, Japanese, Korean, Arabic, Russian, Hindi and Hebrew.

The architecture has one detail worth knowing. Three of every four layers use sliding-window attention with a 4,096-token window and rotary position embeddings; every fourth layer uses global attention with no positional embeddings at all. That interleaving is how a dense model of this size keeps a 256K window affordable to serve, and it is why the long-context behaviour can differ from a model that attends globally everywhere. If you depend on the full window, test retrieval from the middle of long inputs rather than assuming. An advertised window and a usable one are not the same thing.

OpenRouter lists 8,192 maximum output tokens. Short outputs are normal for a model aimed at answering from documents, but it rules Command A out for long-form generation in a single call.

Built for answers that cite their sources

Cohere's whole Command line is shaped around retrieval-augmented generation, and Command A carries that forward. The model card describes training for the final generation step of RAG: you pass document snippets, ideally 100 to 400 words each, and with citations enabled it marks supported spans and points them back to the source document.

That is a genuinely useful property in regulated settings. A cited answer can be audited; an uncited one has to be trusted. It also gives you a cheap hallucination check, because a claim without a citation is a claim the model did not ground. The model is also trained for conversational tool use, with tools described in JSON schema.

Cohere's documentation claims 150 per cent higher throughput than Command R+ 08-2024. That is a vendor figure without published test conditions, and speed on any hosted endpoint depends on the provider and its load, so budget latency from your own measurements rather than the launch claim.

The two-GPU claim, read carefully

Cohere says Command A runs on two A100 or H100 GPUs, against far larger footprints for comparable private deployments. For a dense 111B model this is plausible but tight: in BF16 the weights alone are around 222 GB, more than two 80 GB cards hold, so the two-GPU figure implies quantised weights. That is a reasonable trade for many workloads, but measure quality on the quantisation you will actually serve rather than on Cohere's hosted endpoint.

The broader point stands: a dense model with no expert routing is predictable to serve. Memory and compute scale together, batching behaves the way you expect, and there is no expert load-balancing to tune.

The licence is non-commercial

The Hugging Face release is CC-BY-NC, with Cohere Labs' Acceptable Use Policy on top. You can download the weights, study them, fine-tune them and publish research. You cannot run them in a commercial product without a separate agreement with Cohere, whose commercial route is its API or a private deployment contract.

This is where "open weights" and "open source" part company, and it catches teams who prototype on downloaded weights and only read the licence at launch. Open weights vs open source covers the distinction in general.

Command A+ is the sharp contrast: it ships under Apache 2.0, so the self-hosting story that Command A only offered for research is now available commercially from the same vendor.

Command A or Command A+

On OpenRouter at the time of writing, Command A lists at $2.50 per million input tokens and $10 per million output. Command A+ lists at $0.30 and $1.50, adds image input and reasoning, and covers 48 languages. On price, licence and capability breadth, A+ wins.

Pick Command A when one of these holds:

• You need more than 128K tokens of input. Command A's window is 256K; Command A+ accepts 128K input plus 64K output. For very large document sets in one call, that is the deciding difference.
• You have a validated pipeline. If prompts, citation parsing and evaluation are tuned to Command A, a model change is a regression risk to measure, not a free upgrade.
• You want a dense model's serving profile for a private deployment already licensed through Cohere.

Otherwise, start new work on Command A+.

How to evaluate it for your RAG workload

1. Collect thirty real questions with known answers and the documents that contain them.
2. Run them with citations on, and check two things separately: is the answer correct, and does every factual sentence carry a citation to a document that actually supports it.
3. Repeat with the relevant document placed early, middle and late in a long context, to see whether the interleaved attention loses anything.
4. Run the same set on Command A+ and compare cost per correct, fully cited answer.

Citation precision is the metric that matters most for this family, and it is the one generic leaderboards never report. RAG vs long context covers when to retrieve rather than fill the 256K window at all.