Is it Canada’s turn for a DeepSeek moment?
Canada’s leading large-language model (LLM) developer Cohere has unveiled its new Command A model, which the company claims is faster and uses less computing power than other global competitors.
Nick Frosst
“Command A was in development long before the DeepSeek release, but it certainly validated our approach.”
Cohere
Cohere pitches Command A as a “max performance, minimal compute” solution for its enterprise clients. Last month, tech stocks momentarily crashed over Chinese artificial intelligence (AI) company DeepSeek’s model capabilities, which it claimed to achieve with only a fraction of the funds and compute resources compared to US tech giants.
At the time, Cohere co-founder Nick Frosst told BetaKit that DeepSeek’s model proved their point that “innovation and efficiency, not excessive compute,” is key to AI development.
“Command A was in development long before the DeepSeek release, but it certainly validated our approach,” Frosst said as part of today’s announcement. “We’ve never believed in things like AGI or the bigger is better mentality to building models. We’re instead focused on running a capital-efficient business and building products that solve real-world problems for our customers.”
The company says its new LLM is faster than DeepSeek’s v3 model and OpenAI’s GPT-4o model released in November. It also runs with twice the context length of leading models, allowing it to sift through larger documents. Context length is the amount of information, broken down as tokens, that an LLM can process at the same time.
For comparison, DeepSeek v3 requires at least eight graphics processing units (GPUs) to run with 128k context length. Command A needs only two GPUs with 256k context length.
RELATED: Cohere leaders think DeepSeek proves their point about AI innovation
According to Cohere, Command A also outperforms GPT-4o and DeepSeek v3 on metrics such as inference efficiency (the resource-to-output ratio of generating a response) and certain retrieval-augmented generation (RAG) tasks, which involve a model’s ability to retrieve information from the right sources.
Cohere has never been at the top of the model performance rankings for speed, especially compared to leading LLMs. According to the independent AI model index Artificial Analysis, Cohere’s previous models are not among the top 50 by AI intelligence index. OpenAI models are in the top three, followed by DeepSeek v3 and Anthropic’s Claude 3.7 Sonnet Thinking. However, these rankings are subject to constant change as companies release new models and optimizations.
Though Cohere is one of the most well-capitalized Canadian AI companies, its compute spend also pales in comparison to global competitors. Cohere secured $500 million USD in financing at a $5.5-billion valuation last year ($687 million at $7.6 billion CAD), not to mention the $240 million CAD committed by the federal government towards a Cohere data centre. In comparison, Meta said it would spend up to $65 billion USD on AI infrastructure this year. OpenAI raised $6.6 billion USD in October—more than Cohere’s valuation.
In a blog post, the company noted that Command A’s efficiency is especially key for its enterprise clients, some of whom may be looking to cut costs.
“With these efficiency gains, any business can run AI to increase productivity for employees with agents that can automate work,” Frosst said.
Command A will be “seamlessly integrated” into North, a customizable workplace AI platform the company launched in January. The platform integrates with a company’s in-house applications and allows users to automate complex tasks with AI agents. It also launched a finance-specific edition, North for Banking, in collaboration with the Royal Bank of Canada.
Cohere’s model is also available in 23 languages. The company says it outperforms DeepSeek v3 and GPT-4o in accurately responding to English prompts in Arabic, following the release of its Command R7B Arabic model serving businesses in the Middle East and Northern Africa.
Feature image courtesy Cohere.