What does the Apache 2.0 licence on Command A+ actually permit?

Apache 2.0 is a permissive open-source licence. You can download the weights, self-host them, fine-tune them, and use the model commercially — including inside paid products — without a separate Cohere agreement. It also grants an explicit patent licence. This is broader than typical 'open-weight but restricted' releases that cap commercial use or revenue.

What hardware do I need to run Command A+?

Cohere says Command A+ runs on as few as two H100 GPUs at W4A4 4-bit quantisation. It also ships in BF16 (16-bit) and FP8 (8-bit) for teams that have more memory and want higher-precision inference. The 4-bit tier is what makes a 218B-parameter model affordable to self-host.

How big is Command A+ and how does the Mixture-of-Experts design help?

It is a 218B-parameter sparse Mixture-of-Experts model with 128 experts, but only about 25B parameters are active per token — eight routed experts plus one shared expert. You store a large model but pay the compute cost of a much smaller one at inference, which is why it fits on modest hardware.

Why does Command A+ matter for sovereign or on-prem teams?

Because the weights are downloadable under Apache 2.0, organisations in India and the UK can run the model entirely inside their own infrastructure — no data leaves the building, no usage is metered by a vendor. Combined with 48-language support, 128K context and native citation grounding, that suits regulated, data-residency-sensitive workloads.

Cohere Command A+ Goes Apache 2.0: Open Weights You Can Ship

What changed

The licence is the headline. Command A+ ships under Apache 2.0 — Cohere's first major Command-family model that is genuinely open for commercial use, with no separate Cohere licence required to put it in a product.
It is a big model that behaves like a small one. 218B total parameters, but a sparse Mixture-of-Experts design activates only about 25B per token, so it runs on as few as two H100 GPUs at 4-bit.
It is built for grounded enterprise work. 128K context, 48 languages, and native citation grounding — aimed squarely at on-prem, agentic and long-context document workflows.
The timing is strategic. It lands weeks after Cohere's merger with Aleph Alpha and on the same day as a sovereign-AI agreement with Spain's Indra Group.

Pro tip

"Open weights" and "open source" are not the same promise. Plenty of recent releases let you download weights but then restrict commercial scale, attach acceptable-use riders, or require a vendor agreement above a revenue threshold. Apache 2.0 removes that ambiguity: read the LICENSE file, confirm it is unmodified Apache 2.0, and you can build a paid product on it without phoning anyone's legal team.

What the Apache 2.0 licence actually permits

This is the part that matters most for builders, so it is worth being precise. Apache 2.0 is one of the most permissive widely used open-source licences. For Command A+ it means an organisation can:

Download and self-host the weights — run the model on your own GPUs, in your own data centre or cloud account.
Use it commercially — including inside paid SaaS products, internal tools and customer-facing features — with no revenue cap and no separate Cohere agreement.
Fine-tune and modify it, and redistribute your modified version, provided you keep the licence and attribution notices.
Rely on an explicit patent grant — Apache 2.0 includes a patent licence from the contributor, which matters to enterprise procurement and legal review in a way that many "community" licences do not.

Compare that with the spread of recent "open-weight" releases. The table below is about licence shape, not model quality — it is the question a procurement or platform team asks before anything else.

Release style	Download weights?	Commercial use	Self-host on-prem
Closed API only	No	Metered by vendor	No
Open-weight, restricted licence	Yes	Often capped / conditional	Yes, within terms
Apache 2.0 (Command A+)	Yes	Permitted, no cap	Yes

Watch out

Apache 2.0 covers the weights and code as released. It does not absolve you of responsibility for how the model behaves in production, for the data you fine-tune on, or for the regulatory obligations that attach to your use case — the EU AI Act, the UK's emerging frontier-AI rules, or India's DPDP Act still apply to you as the deployer. Open weights shift control to you; they also shift accountability to you.

Why a 218B model on two H100s is the real story

A permissive licence is only useful if you can afford to run the thing. This is where the architecture does the heavy lifting. Command A+ is a decoder-only transformer built as a sparse Mixture-of-Experts: 128 total experts and a "dropless" token-choice router that selects 8 active experts per token, plus one shared expert applied to every token. So although the model holds 218B parameters, only around 25B are active for any given token.

In plain terms: you pay storage and memory for a large model, but you pay inference compute closer to that of a 25B dense model. That is the trick that lets Cohere quote a footprint of as few as two H100 GPUs at W4A4 4-bit quantisation. The "dropless" router matters here too — token-choice routing without dropping tokens means the model uses its experts efficiently rather than discarding work when a batch is uneven, which keeps throughput predictable under load. The model is available in three precision tiers, and Cohere highlights that the quantisation is effectively lossless — you are not trading quality for the smaller footprint in any obvious way.

Precision tier	Bit width	Trade-off	Indicative GPU footprint
BF16	16-bit	Highest precision, largest memory	Most GPUs of the three tiers
FP8	8-bit	Balanced memory vs precision	Middle of the range
W4A4	4-bit	Smallest footprint, near-lossless per Cohere	As few as 2× H100

For a cost-sensitive team — a startup in Bengaluru, a scale-up in Manchester — the gap between "needs a rack of accelerators" and "fits on a two-GPU box you can rent by the hour" is the difference between self-hosting being a slide in a deck and self-hosting being a line in next month's budget. This is the same downward pressure on the cost of capable open weights that we have been tracking with the DeepSeek V4 Pro open-weight leaderboard run and the wave of Chinese open-weight coding models shipping through May 2026.

From a verified Builder

"The MoE-on-two-GPUs maths is what changes my architecture review. I do not have to choose between a frontier-grade model and keeping client data on our own tin any more — at 4-bit I can do both, and the citation grounding means I can show a regulator where every answer came from."

— Verified Builder · Bengaluru, IN

Native citations and the sovereignty angle

Command A+ ships with native citation grounding — the model can attribute its answers back to the source passages it drew on, rather than you bolting a citation layer on afterwards. For agentic and retrieval-augmented workflows that is more than a nicety: it is the difference between an answer a compliance officer will accept and one they will not.

Put the three properties together — downloadable Apache 2.0 weights, a footprint you can host on-prem, and citations baked into the output — and you have a model positioned for sovereign AI. That word is doing real work right now. UK firms eyeing the public sector, and Indian enterprises under DPDP-era data-residency expectations, increasingly need a model where they can prove that no data left the organisation's own infrastructure. A metered API cannot make that promise; downloadable weights can. The 48-language coverage and 128K context reinforce the same use case: multilingual document processing, long-context reasoning over contracts and filings, and multimodal document workflows are exactly the regulated, high-stakes tasks where teams cannot afford to ship sensitive material off to a third-party endpoint.

This is the same demand signal behind state-level funding moves such as the UK's £500M sovereign AI fund. Cohere is reading the room: on the same day it released Command A+, it signed a memorandum of understanding with Indra Group, Spain's defence prime, to develop sovereign AI for Spain, Canada and Europe — including a defence workstream. And the model landed days after Cohere completed its merger with Aleph Alpha (April 2026), with Cohere shareholders holding roughly 90% of the combined entity valued near $20B and anchored by a ~$600M investment. The licence change is not a one-off; it is part of a pivot towards selling sovereign, deployable AI to governments and regulated industries.

Shipping with open weights? Put it on your Builder profile.

Every article on AI Tech Connect is written for the people actually building. Browse profiles, shortlist who you want to hire or collaborate with — or add your own.

Browse Builders →

What builders should actually do

So you can run it and you are allowed to ship it. What is the sensible move?

If you are cost-sensitive and already paying per-token to a closed API for grounded RAG, price out the 4-bit tier on rented H100s. The break-even against metered inference arrives faster than most teams assume once volume is steady.
If you are sovereignty- or residency-constrained — public-sector work in the UK, regulated finance or health data in India — the on-prem self-host path is the headline benefit. Start with the BF16 or FP8 tier while you validate quality, then quantise down to 4-bit for production economics.
If you are building agents, lean on the native citations and 128K context rather than reinventing a grounding layer. Pair the model with an open agent runtime such as the Cline open-source agent runtime to keep the whole stack inspectable and self-hosted.
Whatever you do, read the actual LICENSE file in the release before you commit architecture to it. Confirm it is unmodified Apache 2.0 and check the model card for any usage notes attached separately from the licence.

A caveat worth stating plainly: Cohere did not publish head-to-head benchmark scores with this release, so we are not going to quote any. Treat the capability claims here as qualitative — the load-bearing facts are the licence, the architecture and the footprint, and those are the things that change what you can build, not just how it scores on a leaderboard.

The bottom line

Command A+ is a 218B sparse-MoE model that you can download, self-host on two H100s at 4-bit, fine-tune, and ship in a paid product — because it is Apache 2.0, not "open-weight with an asterisk". For cost-conscious and sovereignty-conscious teams across India and the UK, that combination is the one that actually moves a self-hosting plan from the whiteboard into the budget. Read the licence, run the maths on the 4-bit tier, and decide whether on-prem just became affordable for your workload.