Cohere Command Models: Command A+ and Command A

Cohere's newest generative model is Command A+, released on 20 May 2026: a mixture-of-experts model with 218 billion total and 25 billion active parameters, 128K tokens of context, 64K tokens of output, and text and image input. Its weights are on Hugging Face under Apache 2.0, and a 4-bit build runs on one B200 or two H100 GPUs. Checked on 31 October 2026.

The Command lineup

Cohere's models page lists the current Command models with their status. Most of the older ones are still live, and the Command R generation from early 2024 has been retired.

Cohere describes Command A+ as "the last model in the Command A family," and as its first mixture-of-experts model. It unifies features that were split across earlier models: reasoning (from Command A Reasoning), vision (from Command A Vision), translation (from Command A Translate) and tool use. The launch blog sets this out in a table.

Source: Cohere's Command A+ launch post. The practical meaning is that if you were running three Command A models for different jobs, Command A+ is meant to replace them with one set of weights.

Maker, licence and open weights

Cohere is a Canadian enterprise AI company, and Cohere Labs is its research arm. The licence story has two chapters, and it matters which one you are in.

Apache 2.0 for Command A+ is a clear break from Cohere's earlier habit of releasing weights under non-commercial terms. It lets you use, modify and redistribute commercially, with a notice and a patent grant. Cohere frames it as "sovereign AI": models you can run, control and adapt yourself. Read your own licence terms for the older Command A weights before you build on them, since the non-commercial clause bars most product uses. We did not find a page that says whether Cohere sells a commercial licence for them, so ask Cohere.

Context window, output and modalities

Command A+ reads text and images and writes text. Cohere lists 128K tokens of input context and 64K of maximum generation. That is smaller than the 256K window of Command A, a trade you should note if you built long-document retrieval on the older model. The 64K output is much larger than Command A's 8K, which suits agent steps that write long tool-call plans or code.

Languages: 48, including all official EU languages, against 23 for Command A. Thinking can be switched on and off, and the docs point to a reasoning guide for token budgets. In the API, the thinking parameter defaults to enabled, thinking={"type": "disabled"} turns it off, and thinking={"token_budget": 500} caps thinking tokens; Cohere advises leaving at least 1K tokens for the response.

Hardware from the model card

This is the clearest hardware table of any model we cover. The Command A+ card lists example minimum GPUs for each quantisation.

Cohere says all three show negligible differences in benchmark quality and performance, and recommends W4A4 for most uses because it has the best speed and latency and the smallest footprint. The model is not a workstation model. A single H100 does not appear in the table, and a consumer 24 GB card cannot hold about 109 GB of 4-bit weights, which is our arithmetic from 218 billion values at half a byte each. Supported frameworks are vLLM and Transformers.

For the economics of renting those GPUs, see Runpod and Lambda and our comparison Runpod versus Lambda. Two H100s for an hour is the entry price of a private deployment.

Choosing a build

The three builds are the same model at different precision, so the choice is about hardware and speed.

Start with W4A4 and test it on your own prompts. If you find a quality gap on a task that matters, move up one build and measure again. Because Cohere's own numbers say the gap is negligible, a failure in W4A4 is more likely to come from your prompt or your serving settings than from the 4-bit weights.

Serving it

The Hugging Face card names vLLM and Transformers as supported frameworks. The usual pattern is to run vLLM as an OpenAI-compatible server and point your tools at it. We show the shape, not a tested recipe. Check the card for the current flags.

vllm serve CohereLabs/command-a-plus-05-2026-w4a4 --tensor-parallel-size 2

The model card's example is vllm serve CohereLabs/command-a-plus-05-2026-bf16 -tp 4 --tool-call-parser cohere_command4 --reasoning-parser cohere_command4 --enable-auto-tool-choice, commented as being for B200 and to adjust -tp for your device, and it asks for vLLM 0.21.0 or later and cohere_melody 0.9.0 or later. Two H100s means tensor parallelism of 2. Our comparison of Ollama and vLLM explains why vLLM is the usual choice for a model this size: it batches requests and spreads a model over several GPUs, while Ollama is built for single-machine convenience.

Moving from Command A

If you already run Command A, here is what changes.

1. Context drops from 256K to 128K. Re-check any retrieval job that stuffs long documents into one request.
2. Output grows from 8K to 64K, so the model can write longer answers, plans and code in one turn.
3. Images now work on the same model, so a separate vision deployment may no longer be needed.
4. Languages rise from 23 to 48.
5. Hardware changes. Command A needed two GPUs by Cohere's account. Command A+ lists one B200 or two H100s at 4-bit, so the fleet size can stay similar while the capability goes up.
6. Licence improves if you used the non-commercial Command A weights: Command A+ is Apache 2.0.

Run your regression set before you cut over. Cohere's own comparison is against Command A Reasoning, so for a non-reasoning workload your own numbers matter more.

What does it cost?

Cohere's pricing page shows the Command A+ card with the price label "Free", which refers to the open licence, not an API rate. The rates it prints are for Model Vault, its managed dedicated platform, and for retrieval models: for example Embed 5 and Rerank 5 instances from $3.00 to $5.00 an hour, and Parse 5 from $4.00 an hour. No Command per-token price appears on that page or in the launch post, so we quote none.

That leaves three routes to run it:

1. Download and self-host. No licence fee. You pay for GPUs.
2. Model Vault. Cohere's dedicated, fully managed platform with no multi-tenancy. We found no Model Vault hourly rate for Command A+ on the pages we read.
3. Cohere API or private deployment. Pricing by agreement for enterprise use.

Cohere also issues free Trial API keys, with rate limits, that cannot be used for production or commercial purposes. A Production key is billed pay-as-you-go.

Vendor-reported results

The launch post reports a few comparisons. They are Cohere's own, measured by Cohere, mostly against its own earlier model.

Source: Cohere's Command A+ blog post, 20 May 2026, vendor-reported. The post compares against Cohere's own prior model and several open-source benchmarks, and shows charts we could not read as numbers. In its text the post reports τ²-Bench Telecom rising from 37% to 85% and Terminal-Bench Hard from 3% to 25% against Command A Reasoning, an Artificial Analysis Intelligence Index of 37, and no external model comparisons. Cohere also reports internal evaluations inside its North product, scored by an LLM judge, which we do not reproduce.

A Terminal-Bench Hard score of 25 percent is a big jump from 3 percent, and it is also not a score that puts Command A+ among the leaders. We do not see a published SWE-bench figure. Treat Command A+ as an enterprise agent and multilingual model first and a coding model second. For coding, see the best open source LLM for coding.

Where it is served

That table is from Cohere's platform notes, which we read today. Bedrock availability for the new model was listed as not available, so AWS Bedrock is not a route for Command A+ yet; we did not check the Bedrock catalogue itself. Because the weights are Apache-licensed, Hugging Face Inference, Replicate or a GPU host such as Modal can serve it too, though we did not check which of them list it. See AI hosting for the wider options.

Coding agents that can use it

Cohere builds an enterprise workspace called North and does not ship a coding agent of its own that we found. For third-party agents, the model is reachable through Cohere's API and through any OpenAI-compatible server such as vLLM if you self-host. We did not check Command A+ support in Cline, OpenCode or Aider. Cohere's docs mention an MCP server for AI client integration of its documentation, which is a docs helper, not a model route. For picking a tool, read choose an AI coding agent.

Privacy and sovereignty

The strongest privacy option is the one Cohere highlights: run the open weights inside your own network, on your own GPUs. Cohere's pitch for Command A+ is "sovereign AI," and its pricing page offers private deployments on premises or in isolated cloud networks. For the hosted API, terms depend on your agreement. Cohere's privacy policy says retention of inputs and outputs on its platform is generally 30 days for enterprise users, that enterprise users can control training on their data, and that it does not train on data from private or third-party cloud deployments.

Limits and gotchas

• Window is smaller than Command A. 128K against 256K.
• Licence split. Apache for Command A+, non-commercial for the older Command A weights.
• Not a workstation model. The smallest listed setup is one B200 or two H100s.
• No public per-token price on the pricing page we read.
• Bedrock gap. Cohere's platform table lists Command A+ as not on Bedrock.
• Coding evidence is thin. The one coding-agent figure we found is 25 percent on Terminal-Bench Hard, from the vendor.
• Trial keys are non-commercial.

Our take

Command A+ is a model for organisations that want control. Apache 2.0, a clear GPU table, 48 languages, vision, and a design that fits on two H100s add up to a good fit for a company that must keep data on its own servers and works in several languages. It is not the model to pick for the best coding results, and it is not the cheapest way to get tokens. If your constraint is "no data leaves the building," compare it with the other open models on the open source LLM page, such as DeepSeek, Qwen and gpt-oss, and run all of them on your own tasks.

Who should pick it

Related pages

• Open source LLM for the licence differences
• Mistral models and Qwen models for other multilingual open-weights families
• Cloud versus self-hosted AI for the cost maths
• Compare AI models side by side for the interactive table