Cohere 2026 Impact Report Intelligence in service of human wellbeing
We believe AI's highest purpose is to enhance human wellbeing. This report shows where that belief has turned into working technology — in hospitals and laboratories, in classrooms and ministries, and in 101 languages the frontier had skipped.
We founded Cohere on a simple conviction: that the value of artificial intelligence is measured by what it does for people, and that the surest way to serve people is to put capable, trustworthy intelligence directly into the hands of the institutions they depend on.
Human wellbeing is not one thing. It is health, and knowledge, and the dignity of being addressed in your own language. It is meaningful work, functioning public services, safety from harm, control over your own information, and a habitable planet to enjoy all of it on. A company that says its highest purpose is wellbeing owes an account across every one of those dimensions — which is how this report is organised.
The account is a good one. Our models are accelerating literature review and scientific analysis in biopharma, where the time between a question and an answer is measured in lives. They are supporting teaching, research and student services across one of the world's great universities, and building the talent pipeline behind it. Through Aya, they speak 101 languages — built with more than three thousand researchers in 119 countries — and through Tiny Aya they now do it on a phone, without a data centre in sight. They serve governments in Canada, the United Kingdom and beyond, so that public institutions can modernise without surrendering control of citizens' data to anyone.
That last point is the thread through all of it. We build intelligence that institutions can own, run and govern themselves — on their own premises, inside their own security perimeter, under their own law. North and Compass are not sold in any configuration where we hold a customer's data. That is a structural choice, not a marketing one, and it is why a hospital, a ministry or a bank can adopt frontier AI at all.
Building capably also means building efficiently. Command A serves at frontier quality on two GPUs where comparable systems needed thirty-two. Command A+ runs on a single accelerator. Tiny Aya runs in a pocket. Every one of those is a reduction in the energy, hardware and cost required to deliver the same benefit — and the reason our research team has argued publicly for standardised ways to measure and compare model efficiency across the whole industry.
This is our first impact report, and we publish it in that spirit: proud of the work, precise about the evidence, and clear about where we are deepening the measurement next. The final chapter sets out exactly how, over the coming twelve months.
Illustrative leadership message · Prepared for a first reporting cycle
101 languages in the Aya family, built with 3,000+ researchers across 119 countries. Frontier weights released under Apache 2.0. Models small enough to run in a clinic, a ministry or a pocket. Capability that cannot reach people is not capability.
On-premises, isolated VPC and sovereign deployment, backed by SOC 2 Type II, ISO 27001, ISO 42001 and FedRAMP High. When a hospital or a government adopts our technology, the data, the model and the off switch all stay theirs.
Two GPUs where comparable models needed thirty-two. A 218-billion-parameter model served from a single accelerator. Seventy languages on a phone. Right-sizing is how the same benefit reaches more people for less energy.
Wellbeing is multidimensional, so our reporting is too. Below: the ten domains this report covers, and the strength of the evidence behind our contribution in each — from live, verifiable programmes through to the areas where measurement is still being built.
Evidence behind our contribution
64/100 average evidence strength
Each domain scored 0–5 on the strength of publicly verifiable evidence: a live programme, published artefacts, independent validation, named partners, and measured outcomes. Domains where Cohere already operates at scale with published research and named institutional partners score highest. Domains marked Building have active programmes with outcome measurement still being instrumented — the subject of chapter 13.
Access delivered
Of roughly 7,000 languages spoken worldwide, only about 1,500 have readily available digital data. Aya was built to close that gap deliberately rather than incidentally, and Tiny Aya now carries 70+ of those languages on consumer hardware — putting capability where compute is scarcest.
Efficiency as a wellbeing contribution
Every reduction in the hardware needed to serve a model is a reduction in cost, energy and emissions per unit of benefit delivered — and an expansion in who can afford to deliver it. A hospital that can run frontier AI on two GPUs can run it at all; a developer with one accelerator can build for a language no hyperscaler serves. Efficiency is not only an environmental measure for us. It is the mechanism by which capability reaches people.
Current Command A, Command A+, North Mini Code and Cohere Transcribe all released with open weights under Apache 2.0.
Three domains where the contribution is already concrete, named and verifiable — the strongest evidence in this report that intelligence is reaching people rather than only enterprises.
→ built on the Reliant AI acquisition, May 2026
An agentic system for R&D, clinical development and scientific analytics. Reliant's technology automates literature review, systematic reviews and scientific data extraction — work that has historically consumed months of researcher time between a question and an answer.
→ University of Toronto and University of Waterloo, 2026
North serves as an orchestration layer across teaching, learning, research, student services and administration at the University of Toronto, supporting its AI Kitchen for privacy-conscious evaluation. The Waterloo partnership builds Canada's AI talent pipeline from Autumn 2026.
→ Tiny Aya, 3.35B parameters, four regional variants
Global, Earth, Fire and Water — variants weighted toward Africa and West Asia, South Asia, and Asia-Pacific and Europe. Expedition Aya put them into the hands of builders worldwide, producing tools for education, accessibility and local deployment.
Cohere Labs' AI Language Gap primer sets out the problem these releases address: Dutch, with 29 million speakers, has roughly two million Wikipedia entries; Somali, with 18 million speakers, has around 5,500. Language coverage in AI has tracked digital wealth rather than human need. The Aya model paper received the Best Paper Award at ACL 2024.
Strong programmes deserve strong measurement. Over the coming year we are instrumenting the outcomes our technology produces, extending our environmental accounting, and bringing all of it to a standard that can carry external assurance. Chapter 13 sets out the plan in full.
Illustrative model
Model efficiency is the lever with the widest reach: it lowers the energy, cost and hardware needed for every inference served, everywhere, at once. This transparent sketch uses the sector inference intensity published in our own Efficient AI primer to show how the arithmetic scales. Move the sliders.
Estimated An illustrative calculation using published sector figures, shown in full so the method can be checked. Cohere's own per-model energy intensity is being instrumented and will be published in the next cycle.
Report architecture
Chapters 1 and 2 set out our wellbeing framework and how we report against it. Chapters 3 to 12 take each dimension of wellbeing in turn — health, knowledge, language, work, public life, autonomy, safety, the environment, our own people, and the communities we work in. Chapter 13 is how we deepen the measurement.
Read on
Cohere · Illustrative first impact report
Reporting period: calendar year 2025, with selected developments through August 2026. Organised around the ten dimensions of human wellbeing our work touches. Every claim carries a status label so readers can see what is live today, what is being measured, and what comes next.