Command A+

What it is

Command A+ is a large language model that Cohere released on 20 May 2026 as an open-source mixture-of-experts model under the Apache 2.0 licence^[2][4]. The company positioned it for sovereign and critical-infrastructure deployments^[2][4] — settings where an organisation wants a capable model it can operate itself rather than only call remotely. Cohere says the model is built to remain fit to run entirely within customer environments^[5], and describes it as advancing its mission to make sovereign AI a technological reality^[5]. Cohere's own launch post is dated 20 May 2026 and announces the model as open-source from that day^[5].

How it is built

Command A+ is a sparse mixture-of-experts (MoE) Transformer with 218 billion total parameters and 25 billion active parameters^[1]. The sparsity comes from routing: of its 128 experts, 8 are active for each token, plus one shared expert that is always in play^[1]. In a mixture-of-experts model each token is sent to a small subset of specialised sub-networks rather than through the whole network, so the compute spent per token stays well below what a dense model of the same total size would need. Cohere makes that contrast directly with its own earlier model, describing Command A+ as 218B total and 25B active against the 111B dense architecture of Command A Reasoning^[5].

Its attention layers follow a 3:1 ratio, an arrangement first introduced in Command A^[1]. The model accepts up to 128k tokens of input context and can generate up to 64k tokens of output^[1]. It takes text and images as input and returns text, and it can reason, writing its thinking between dedicated thinking tags^[1]. It supports 48 languages^[1].

What it takes to run

The model card sets out minimum hardware configurations for self-hosting. In BF16 the floor is 4x B200 or 8x H100; in FP8 it is 2x B200 or 4x H100; with the recommended W4A4 quantisation a single B200 or 2x H100 is enough^[1]. Quantisation here means storing and computing with lower-precision numbers so the model fits into less memory, which is what allows the same weights to run on a much smaller machine. The published hardware table matters for the deployment story: a model that can be served on a handful of accelerators is one an organisation can plausibly keep inside its own infrastructure.

Licensing

Command A+ is Cohere's first Command model fully licensed under Apache 2.0^[1][2]. Earlier Command models — Command R, R+ and A — shipped under CC-BY-NC weights that restrict commercial use^[1][2]. Moving to a permissive open-source licence changes what downstream users may do with the weights: they can be run, modified and built into commercial products without the non-commercial restriction. Cohere frames the release around control of deployment rather than around benchmark competition, saying the model is built to remain fit to run entirely within customer environments^[5].

Where it sits

Cohere's current line-up places Command A+ as a flagship model, listed on Wikipedia alongside Command A Reasoning, Vision and Translate^[4]. Hosted access is available too: OpenRouter lists Cohere's price at $0.30 per million input tokens and $1.50 per million output tokens^[3], so teams that do not want to run the weights themselves can pay per token instead. The combination is unusual in one respect — the same model is offered both as Apache 2.0 weights for local deployment and as a metered hosted endpoint. For Cohere, whose pitch centres on enterprises and governments keeping data and models under their own control, Command A+ is the model that carries that argument into open weights.