Cohere has unveiled its latest AI model, Command A, offering a solution that combines high performance with remarkable efficiency. Unlike traditional models, such as GPT-4o and DeepSeek-V3, which require many GPUs to operate at full capacity, Command A runs on just two, significantly reducing energy consumption while maintaining top-tier performance.
According to Cohere, Command A is “an auto-regressive language model that uses an optimized transformer architecture.” Like other models from Cohere, Command A has been trained specifically for tasks like the final step of Retrieval Augmented Generation (RAG).
“After pretraining, this model uses supervised fine-tuning (SFT) and preference training to align model behavior to human preferences for helpfulness and safety. The model features three layers with sliding window attention (window size 4096) and RoPE for efficient local context modeling and relative positional encoding. A fourth layer uses global attention without positional embeddings, enabling unrestricted token interactions across the entire sequence.”
The model has been trained on 23 languages: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, Chinese, Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian. It uses 111 billion parameters and provides a context window of 256K token.
Command A’s energy efficiency sets it apart as a compelling choice for businesses looking to leverage AI without incurring the typically high operational costs associated with more resource-intensive models.
Cohere’s decision to prioritize energy conservation in its design not only speaks to growing concerns about the environmental impact of AI but also positions the company as a leader in the rapidly evolving field of green AI.
As the global demand for AI continues to grow, businesses are increasingly looking for AI tools that meet performance needs without the heavy environmental cost, and Command A answers that call. The AI market is dominated by models that rely heavily on computational power, but Command A redefines this standard.
The model’s reduced resource consumption makes it a highly scalable solution for enterprises across various sectors, where sustainability is increasingly becoming a key business priority.
Command A Performance vs GPT-4o and DeepSeek V3: Benchmark Results and Human Evaluations
Cohere’s Command A has demonstrated impressive performance across various benchmarks and evaluations, often outperforming both GPT-4o and DeepSeek-V3 in several key areas.
In the Human Preference Evaluation, Command A performs strongly across multiple domains. In General Business, it slightly surpasses GPT-4o with 50.4% compared to 49.6%. In STEM, Command A holds an edge with 51.4% over GPT-4o’s 48.6%, while in coding, GPT-4o leads with 53.2% compared to Command A’s 46.8%. This demonstrates Command A’s broad applicability across diverse fields, including business and technical tasks.
In terms of Inference Efficiency, Command A outperforms both GPT-4o and DeepSeek-V3. It generates 156 tokens per second at 1K context, far surpassing GPT-4o at 89 tokens and DeepSeek-V3 at 64 tokens. This makes Command A more efficient in terms of processing power, enabling faster response times and handling larger volumes of data with greater ease.
Command A also excels in real-world benchmarking tests. In tests like MMLU, Taubench, and SQL, Command A consistently ranks alongside or above GPT-4o, outperforming DeepSeek-V3 in coding tasks like MBPPPlus and RepoQA.
This places Command A in a competitive position in both academic and business-related applications, confirming its robustness in handling complex tasks.
In terms of Arabic Crosslingual Language Accuracy, Command A leads with an impressive 98.2% accuracy, surpassing DeepSeek-V3 at 94.9% and GPT-4o (Nov) at 92.2%. This achievement highlights Command A’s superior ability to handle complex English instructions in Arabic, which is particularly important for global applications that require multilingual support.
Moreover, Command A also excels in ADI2 score (ability to respond in the same Arabic dialect as the prompt), achieving 24.7, significantly outpacing DeepSeek-V3 with 15.7 and GPT-4o at 15.9. This makes Command A a highly effective model for dialect-specific tasks, catering to a more diverse range of Arabic dialects.
Finally, in multilingual human evaluations, Command A shows strong performance in languages such as Arabic, Portuguese, and Spanish, winning a significant percentage of the evaluations. Its performance in Arabic is particularly notable, where Command A outshines DeepSeek-V3, further solidifying its competitive advantage in multilingual environments.
With its superior performance in inference efficiency, multilingual accuracy, and coding benchmarks, Command A emerges as a highly capable model that is well-suited for a wide range of applications, from business processes to technical problem-solving. Its efficiency and accuracy make it a compelling choice for enterprises seeking AI solutions that do not compromise on quality or performance.
Command A’s Integration with Cohere’s Broader Strategy
The introduction of Command A fits within a broader vision for Cohere to provide businesses with a comprehensive suite of customizable AI tools.
This vision is evident in Cohere’s North platform, which was launched in January. The North platform is designed to integrate Command A’s efficiency with the automation of core business functions, such as document analysis, customer service automation, and HR tasks.
By offering flexible, scalable AI solutions, North is not just a product, but a key piece of Cohere’s enterprise AI ecosystem that helps businesses reduce costs while increasing operational efficiency. What sets North apart is its ability to integrate Command A’s low-resource architecture into business workflows, making it suitable for healthcare, finance, and manufacturing—sectors where operational cost control and security are paramount.
The platform’s ability to ensure data privacy while handling AI-driven tasks gives it a competitive edge, particularly for businesses that must comply with stringent regulatory standards.
As Cohere’s AI offerings evolve, Command A stands out as an essential model within their portfolio. Its integration into North enhances the platform’s ability to meet the growing demands of businesses looking for reliable AI solutions with a low energy footprint.
In addition, Aya Vision, launched in March 2025, represents another example of Cohere’s broader strategy to offer open-weight AI solutions. Aya Vision’s multimodal capabilities and open-weight design align with Cohere’s push for transparency and customizability in AI, ensuring that developers and businesses alike can adapt it to their specific needs.
Legal Challenges: Copyright and Data Use in AI
Despite the technological advancements by Command A and other products, Cohere faces significant legal hurdles.
In February 2025, a lawsuit was filed by major publishers, including Condé Nast and McClatchy, accusing Cohere of using their copyrighted content without permission to train its AI models, including the Command family of models.
The plaintiffs argue that Cohere’s use of retrieval-augmented generation (RAG) technology to enhance the model’s performance involves replicating their content without sufficient transformation or authorization.
Cohere has responded by defending its use of RAG as being within the bounds of fair use, but the issue remains contentious. The lawsuit highlights the complex challenges faced by AI companies regarding data usage and intellectual property rights.
As AI models become more integrated into business workflows, the question of whether it is ethical—or legal—to train AI models on publicly available content without explicit permission will be crucial in shaping the future of AI technologies.
The outcome of this lawsuit could have far-reaching consequences not only for Cohere but for the entire AI industry, potentially setting new precedents for how AI models are trained in the future.
As AI continues to be integrated into every facet of enterprise business, questions surrounding data ownership and AI-generated content will become increasingly important, especially in the context of open-weight models like those Cohere is advocating for.
Cohere’s Position in the AI Market
Despite the notable advantages of Command A and Aya Vision, Cohere is not without competition. Proprietary models like OpenAI’s GPT-4o and Google’s Gemini remain dominant players in the market, offering unmatched performance but at the cost of high resource consumption and restrictive access.
These models cater to large-scale enterprises willing to invest heavily in AI infrastructure, but the closed-source nature of these models limits flexibility and customization options. Cohere’s open-weight approach offers a distinct alternative.
The company’s decision to focus on open-access AI models, such as Aya Vision, provides significant flexibility, enabling developers to fine-tune models for specific tasks and industries.
This is a major advantage for researchers, startups, and small businesses, which may lack the resources to navigate the complex licensing agreements tied to proprietary models. As AI becomes more integral to business operations, Cohere is positioning itself as a key player in the movement toward open-source AI.
Furthermore, Cohere’s ability to offer energy-efficient models with top-tier performance gives it a competitive edge over other players in the market. While OpenAI and Google have long been the industry standard, Cohere’s Command A offers an attractive alternative for businesses seeking AI solutions that don’t come with the hefty energy costs associated with more traditional systems.
