Downloadable weights
The model can run without depending exclusively on a Cohere-hosted inference endpoint.
Canadian AI models
Model origin matters. But it is only one layer of sovereignty.
Canada has its own foundation-model ecosystem, including enterprise AI company Cohere. A Canadian model can reduce certain external dependencies, but it still has to run somewhere, process data somewhere and operate inside an architecture someone controls.
Underlabs evaluates the complete system, not only the logo on the model.
The model inside the system
An LLM matters. It is not the whole system.
A large language model interprets language, reasons over information and generates responses. In an enterprise system, it usually works with an application, company data, tools and business systems.
Sovereignty depends on that complete path, not only the model name.
- 01User
- 02Application
- 03Company data and retrieval
- 04Language model
- 05Tools and business systems
- 06Response or action
Real Canadian capability
Canada also develops foundation models.
Canada is not only a consumer of foreign AI technology. Its current national strategy treats foundation models, sovereign compute and domestic infrastructure as strategic capabilities. It identifies Cohere, founded in Toronto in 2019, as a Canadian frontier-model company focused on enterprise and government uses.
That context validates a domestic commercial capability. It does not mean a Canadian model is automatically right for every system.
Cohere as a case study
Command A+ shows the architectural value of open weights.
As of September 2026, official documentation describes Command A+ as a sparse mixture-of-experts model for enterprise workloads, with reasoning, tools, structured outputs, text and image input, and multilingual support.
Permissive license
It is released under Apache 2.0, providing substantial deployment flexibility. This is not legal advice.
Efficient private deployment
Cohere documents a minimum W4A4 configuration of one B200 or two H100s, a relatively contained footprint for a model at this scale.
Enterprise capabilities
Reasoning, tools, structured outputs, multilingual operation and a 128,000-token context window are relevant to business systems.
Verified model facts: released May 20, 2026; 218B total and approximately 25B active parameters; text output; 64,000-token maximum output; 48 languages, including French. Official Command A+ documentation.
Precise terminology
Open weight does not always mean open source.
Open weights
The trained model parameters can be downloaded and run independently.
Open source
Depending on context, the term may imply broader openness across code, weights, training methods, datasets or licensing.
Cohere describes Command A+ as open source and releases it under Apache 2.0. Underlabs still uses the more technically conservative term “open-weight model” when discussing downloadable, independently deployable parameters.
Origin and deployment
Canadian model ≠ Canadian AI system.
Model nationality and control over inference are separate decisions.
OriginDeployment
Externally run
Privately run in Canada
Canadian model
Canadian model, external inference
Canadian origin, but infrastructure and inference remain externally managed.
Canadian model, private Canadian compute
Can combine Canadian development, residency, private inference and infrastructure control.
International model
International model, external API
Often the easiest deployment with excellent capability, but externally controlled inference.
International open-weight model, private Canadian compute
Non-Canadian origin, with Canadian residency and customer-controlled inference.
Sovereignty is architectural. Model nationality is one variable inside that architecture.
Five different questions
Do not confuse model nationality with data nationality.
These distinctions make requirements testable and prevent a provider choice from becoming a sovereignty promise.
Model origin
Where was the model developed?
Data residency
Where does company information live?
Inference location
Where does the model actually run?
Infrastructure control
Who controls the computers?
Operational control
Who can access, change or disable the system?
These are five different questions.
When origin matters
An additional consideration, not an automatic answer.
Canadian origin can have strategic, operational or procurement value when the model’s capability first satisfies the workload.
Government and public sector
Domestic capability may matter when procurement or strategic autonomy requires it, without making Canadian models mandatory.
Sensitive enterprise systems
Some organizations want fewer foreign dependencies in critical systems.
Regulated industries
Deployment flexibility and data control may matter more than model nationality itself.
Long-term independence
Downloadable weights can allow a model to keep operating without exclusive dependence on a public API.
Bilingual requirements
French and English performance should be measured directly on real tasks.
Canadian technology
An organization may prefer Canadian-developed technology when its capability satisfies the workload.
The best model for the work
An international model may still be the right answer.
OpenAI, Anthropic, Google or another provider may deliver materially better results for a workload: reasoning, coding, multimodality, context, agents, cost, latency, tooling or ecosystem.
If sovereignty requirements permit external inference, that may be the correct architecture. Underlabs does not sacrifice system quality for model nationality alone unless the client’s requirements justify that trade-off.
- Capability
- Reasoning quality
- Task accuracy
- Tool use
- Languages
- Latency
- Context
- Privacy
- Licensing
- Deployment
- Hardware
- Cost
Empirical model selection
Benchmark models on the customer’s real work.
General leaderboards provide signals. They do not necessarily measure quality on the documents, tools, formats, vocabulary and failure modes of the actual system.
Underlabs compares accuracy, hallucinations, reasoning, French and English, structured outputs, tool use, latency, cost, hardware, privacy and deployment flexibility.
Model AModel BCanadian modelPrivate model
Compare results and constraintsIntentional portability
The best model can change.
Today’s strongest model may not remain the strongest six months from now. An Underlabs service layer can reduce direct coupling between the application and one provider.
A swap is never guaranteed to be simple. APIs, tool-calling formats, prompts, context limits and structured outputs differ. Abstraction reduces unnecessary lock-in; it does not remove integration and evaluation work.
ApplicationUnderlabs AI layer
CohereExternal modelPrivate model
More than language models
An architecture can combine several specialist models.
Embeddings, rerankers, vision, transcription, translation, classification, OCR, code and other specialist models may serve a particular step better than the main LLM.
Small private model+retrieval+business rules+tools
For a bounded workload, that combination can produce a better business system through more predictable cost, latency, infrastructure and control. This is not a general claim that smaller models outperform frontier models.
English and French
Language support must be tested, not assumed.
Command A+ documents French among its 48 languages. That does not automatically establish equal performance on Quebec business terminology, Canadian government language, customer-service language, informal Quebec French or a specialist domain.
Same real taskEnglishFrançaisCanadian French, where relevantCompare results
Underlabs model-selection process
Choose with evidence, then keep the architecture adaptable.
The durable capability is model selection, evaluation, integration and private deployment, not dependence on Cohere or any other provider.
- 01
Define the workload
Specify the required outcome, data, tools and cost of an error.
- 02
Identify sovereignty requirements
Decide what must remain in Canada and what may use external services.
- 03
Shortlist models
Compare Canadian and international, private and managed, large and small options.
- 04
Benchmark real tasks
Measure quality, errors, languages, structured outputs, tool use and latency.
- 05
Evaluate infrastructure
Confirm that the model can realistically run in the required environment.
- 06
Compare full economics
Include inference, GPUs, operations, engineering and scaling.
- 07
Integrate intentionally
Reduce unnecessary lock-in without pretending every model is interchangeable.
- 08
Re-evaluate
Repeat evaluation as models and requirements change.
From model to compute
A privately deployable model still has to run somewhere.
Model selection leads to the next decision: GPU capacity, Canadian cloud, private data centres, storage, networking and inference servers. The model cannot be selected independently of the available compute environment.
Fit the model to the workload
Benchmark the models that actually satisfy your requirements.
Share the tasks, data, languages, tools, infrastructure constraints and sovereignty boundary. Underlabs can evaluate Canadian and international options without prescribing one provider.