Changelog
May 25, 2026
Cohere Just Open-Sourced a Frontier Model You Can Own
On May 20, 2026, Cohere did something the big AI labs have spent three years avoiding. It took its strongest model, wrapped it in the most permissive license in software, and handed it to anyone with two GPUs.
The model is Command A+. It is the first time Cohere has released a frontier-grade model under Apache 2.0 — meaning you can download it, run it on your own hardware, modify it, and build a commercial product on top of it without asking permission or paying a per-token toll. No usage caps. No vendor lock. No data leaving your walls.
If that sounds like a technical footnote, it is not. It is a shift in where the money and the leverage live. Here is the plain-English version of what happened and why it matters for anyone running a business on AI.
What Cohere Actually Shipped
Command A+ is a 218-billion-parameter model. That number sounds enormous, and it is — but the clever part is that it only uses 25 billion of those parameters for any given word it generates. The rest sit idle until needed.
The hard facts, verified against Cohere's own release: a 128,000-token context window, 48 supported languages, native handling of text and images, and built-in citations so the model shows its sources. On agentic coding tasks it jumped from 3 percent to 25 percent on Terminal-Bench Hard. On a telecom reasoning benchmark it went from 37 percent to 85 percent. It scored a 37 on the Artificial Analysis Intelligence Index, putting it at the front of the open-weight pack.
Translation: this is not a toy. It is a serious, fast, multilingual, multimodal model that happens to be free to own.
Why a 218B Model Runs on Two GPUs
The reason a 218B model runs on as few as two GPUs comes down to its architecture — a "mixture of experts." Instead of one giant brain firing every neuron for every task, the model is built from 128 smaller specialists. For each word it generates, a router picks the 8 most relevant experts plus one generalist that always runs. The other 119 stay dark.
You get the knowledge of a 218B model at the running cost of a 25B one. Cohere then squeezed it further with 4-bit quantization that they claim is near-lossless — shrinking the memory footprint so the whole thing fits on two H100s or a single Blackwell chip, generating around 375 words per second with a first response in roughly a tenth of a second.
The headline most people miss: frontier capability no longer requires a data center. It requires a server closet.
What Apache 2.0 Really Unlocks
"Open source" gets thrown around loosely in AI. Most "open" model releases come with non-commercial licenses or research-only terms — you can look, but you cannot build a business. Apache 2.0 is different. It is the same license that runs much of the internet's plumbing. It says: use this commercially, modify it, ship it inside your product, and you owe no one anything.
Cohere is betting on what it calls "sovereign AI" — the idea that enterprises, governments, and developers should be able to run, control, and adapt frontier models entirely inside their own secure environment. As the company put it, empowering engineers with models they can run, control, and adapt themselves is "the most acute challenge facing this generation of AI."
For a hospital, a bank, or a government that cannot send data to a third-party cloud, this is the unlock. The model comes to the data, not the other way around.
The Leverage Just Moved Down the Stack
Here is the part worth sitting with. For three years the assumption has been that the advantage in AI belongs to whoever owns the biggest model and the most cloud infrastructure. When a frontier model becomes a free download that runs on hardware you can buy, that assumption breaks.
The model is no longer the moat. The infrastructure is no longer the moat. What is left — the only thing that is actually scarce — is the workflow. Knowing which problem to point the model at, how to wire it into a real operation, what to automate, and how to package it so a customer will pay for the outcome.
The question quietly stops being "who has the best model" and becomes "who can design and sell the system around it." That is a far better question for an operator than for a trillion-dollar lab. The leverage just moved down the stack, toward the people who build things that work.
Own the Model, Sell the Workflow
At CYSTEMS this is the thesis we already run on. We do not sell access to a model — anyone can rent that now, or as of last week, own it outright. We sell the system: the audit that finds where AI actually creates leverage in your business, and the automation that captures it.
Command A+ makes that case louder, not quieter. As frontier models commoditize, the value migrates to whoever can turn raw capability into a working, owned, defensible operation. Owning the model is now table stakes. Owning the workflow is the game.
If you want to know where AI actually moves the needle in your business — not the hype, the leverage — that is exactly what our audit is built to find.
Changelog