Canadian AI: Cohere Command A+ and the Builders Beyond Chat

Canadian AI is easier to understand when we stop treating it as a single contest to build the most famous chatbot. The ecosystem includes language models, image generation, autonomous transportation, industrial robotics, research institutes, and the infrastructure that helps turn experiments into products. Those categories solve different problems and should not be squeezed into one leaderboard.

Cohere is a useful place to begin. Founded in Toronto in 2019, it now describes a business with offices across Canada and other countries. That Canadian origin is meaningful, but it does not mean every deployment, employee, or customer environment is located in Canada. Command is Cohere's model family, not a separate company. North is an enterprise AI workspace, not another model competing with Command.^[1]

For a business choosing AI, these distinctions are practical. You purchase or deploy a particular capability under particular terms. A country's research reputation can help you discover suppliers. It cannot replace a review of the model version, hosting arrangement, license, integration requirements, and operating costs.

Click to expand

Command A+ is the current story, not an older license

Cohere released Command A+ on May 20, 2026. Its technical announcement identifies the model as command-a-plus-05-2026 and describes a sparse mixture-of-experts architecture with 218 billion total parameters and 25 billion active parameters. The snapshot lists 128K input context, up to 64K generation, support for 48 languages, image and text inputs, reasoning, and tool use.^[2]

Crucially, the current release is available under Apache 2.0. That is not the older noncommercial description sometimes associated with other Cohere releases. Public Hugging Face metadata for CohereLabs/command-a-plus-05-2026-bf16 also identifies Apache 2.0. Teams still need to read the relevant license and deployment terms, but they should evaluate the release in front of them rather than reuse a stale assumption.^[2]

The hardware statement also needs its qualifier. Cohere specifies a minimum of one B200 or two H100 GPUs at W4A4 quantization. That is not a claim that the full-precision model fits on an ordinary laptop. Quantization, serving framework, concurrent users, context length, and workload all affect the resources needed in practice. A minimum configuration is a starting point for capacity planning, not a service-level guarantee.

Cohere reports improvements over earlier Command models on several evaluations. We regard those vendor results as useful evidence for deciding what to test, not independent proof that every customer will see the same gains. The meaningful question is whether the model handles your documents, permissions, and failure cases with acceptable cost and latency.

Choose capability before choosing a flag

North helps explain why a model specification is only part of an enterprise purchase. Cohere describes it as a platform for workplace productivity, with retrieval and agentic capabilities grounded in enterprise data. Command A+ is a component that can support such workflows; the workspace provides a different layer of the product.^[1][2]

Cohere's model catalogue makes the wider stack clearer. Command models cover generation and tool-oriented tasks. Embed supports search and classification representations. Rerank reorganizes search results. Parse extracts structured information from documents, while Transcribe addresses speech recognition. A retrieval product may need several of these capabilities rather than one large model asked to do everything.^[3]

That creates a useful buying discipline. If the problem is finding the right paragraph, improve retrieval before paying for longer generated answers. If the problem is extracting fields from scanned forms, document parsing may deserve its own evaluation. If the problem is performing a business action, the decisive issues may be permissions and confirmation rather than eloquence.

Intueo's editorial view is that this modular perspective is a strength. It encourages teams to connect a capability to a measurable task. It also makes a pilot easier to diagnose: poor search, incorrect extraction, and unsafe tool permissions are different failures with different remedies. Calling all of them an AI problem obscures the work required.

Tiny Aya shows why small and open are separate questions

Tiny Aya offers a different scale of multilingual work. Cohere's catalogue describes a compact 3.35-billion-parameter model family supporting 70 languages, with globally balanced and regional instruction-tuned variants. Its listed context and maximum output are 8K. That makes it worth examining for bounded multilingual tasks where a much larger model may be unnecessary.^[3]

However, open weights do not automatically mean unrestricted commercial use. Public metadata for tiny-aya-global identifies CC BY-NC 4.0 and its gated form explicitly requires agreement to noncommercial use. That is materially different from Command A+'s Apache 2.0 release. A business should investigate separate commercial authorization or suitable API terms rather than assume downloadable weights grant permission for a paid product.^[4]

English and French support can be commercially useful in Canada, and broader language coverage matters for organizations serving diverse communities. Coverage counts are not quality guarantees. A bilingual pilot should include terminology, regional phrasing, mixed-language requests, and examples where the model must preserve a name or legal reference exactly. Sensitive translation still needs appropriate human review.

Beyond language: creative and physical AI

Ideogram's founding announcement describes a company building generative AI tools for creativity, with a team drawing on work in diffusion models, text-to-image generation, and related research, and recruiting in Toronto. It belongs in a discussion of Canadian AI because image creation and visual communication are substantial capabilities in their own right, not side features of office chat.^[5]

Waabi's current public site presents its Physical AI Platform around autonomous transportation, including trucks and robotaxis. Those systems interact with roads, vehicles, and changing environments. Their evaluation necessarily involves physical behavior and operational conditions, not only whether an answer sounds convincing.^[6]

Sanctuary AI describes physical AI for industrial work across robotic hardware, with an emphasis on dexterity, dynamic tasks, and an embodiment roadmap. Its public positioning is about automating work in the physical world. It should not be described as just another general chat-model provider.^[7]

These company descriptions indicate focus, not independent certification of every performance claim on their sites. We are not ranking their safety, commercial maturity, or technical superiority. The useful point is category diversity: creative production, enterprise knowledge work, transportation, and manipulation all present distinct opportunities for businesses. A national ecosystem can be valuable precisely because it contains more than one kind of intelligence product.

Research strength needs a commercialization bridge

Research institutes occupy another role. Mila's public site describes a Quebec machine-learning research community and university affiliations. Vector Institute describes an independent not-for-profit research organization connecting research, talent, and industry adoption. They are not interchangeable with commercial model vendors, even though their partnerships and talent networks contribute to the ecosystem.

For businesses, the attractive combination is access to technical knowledge and a route to operational use. Researchers can help advance methods; vendors and internal engineering teams still need to package those methods into supported products. Customer discovery, distribution, procurement, and implementation are not automatic consequences of publishing excellent research.

Compute is another part of that bridge. Canada's Sovereign AI Compute Strategy announced up to $2 billion across domestic commercial capacity, public infrastructure, and an access fund. The official strategy identifies allocations of up to $700 million, $1 billion, and $300 million respectively, addressing both infrastructure and the cost of access.^[8]

These are policy commitments and program structures, not a guarantee that every firm can immediately obtain funding or that all announced infrastructure is already available. Businesses should check current eligibility and delivery status. The underlying commercial lesson is sound: access to reliable compute can determine whether a promising prototype becomes a service that customers can afford.

A Canadian supplier is not a compliance certificate

Canadian origin can be a reason to engage a supplier. It is not proof of Canadian data residency, regulatory compliance, or complete control over the system. A Canadian-founded company may offer global cloud services; a privately hosted model may still call external tools. The relevant questions concern the actual contract and architecture, including processing location, logs, backups, support access, and retention.

A sensible pilot starts with one recurring workflow and representative examples. Record quality, cost, latency, and reviewer effort against the existing process. Keep early integrations read-only where possible. Before allowing actions, define exactly which tools can be called, what requires approval, and how failures are reversed or escalated.

Canada's opportunity is not to win a nationality contest. It is to translate research, multilingual needs, creative capability, and industrial expertise into useful products. Cohere's Command A+ gives that discussion a concrete, current model release. The broader ecosystem gives it range. The next step for a buyer is to choose the task, verify the terms, and measure whether the resulting system genuinely improves the work.

Specifications and product positioning are attributed to their providers. Deployment recommendations and ecosystem interpretation are Intueo editorial opinion. Public model metadata was checked on October 6, 2026; vendor benchmarks were not independently reproduced.