Cohere has released Command A, a high-performance AI model featuring 111 billion parameters, a 256K context length, and support for 23 languages, on March 16, 2025. The model is designed for enterprise applications, promising a 50% reduction in operational costs compared to existing API-based models.
Meet Cohere Command A
Command A addresses the significant challenges posed by training and deploying large-scale AI models that often require extensive computational resources. Typical models, such as GPT-4o and DeepSeek-V3, demand up to 32 GPUs and extensive infrastructure, which poses a barrier for smaller enterprises. Command A, however, operates efficiently on just two GPUs while maintaining competitive performance levels.
The core architecture of Command A employs an optimized transformer design featuring three layers of sliding window attention, each with a window size of 4096 tokens. This structure enhances local context modeling, allowing the model to effectively manage detailed information across lengthy text inputs. Additionally, it includes a fourth layer that consists of global attention mechanisms, facilitating unrestricted token interactions throughout the entire sequence, thereby enriching its contextual understanding.
Command A achieves a token generation rate of 156 tokens per second, which is 1.75 times faster than GPT-4o and 2.4 times faster than DeepSeek-V3. Its performance in handling instruction-following tasks, SQL queries, and retrieval-augmented generation (RAG) applications has shown exceptional accuracy in real-world evaluations, outperforming its competitors in multilingual scenarios.
Baidu just made AI cheaper: Ernie 4.5 costs 1% of GPT-4.5
The model’s multilingual capabilities extend beyond basic translation, exhibiting superior proficiency in various dialects, including Arabic, with evaluations showing enhanced contextually appropriate responses for regional dialects such as Egyptian, Saudi, Syrian, and Moroccan Arabic. This linguistic versatility is particularly beneficial for businesses operating in diverse language environments.
Performance evaluations indicate that Command A consistently outperforms its peers in fluency, faithfulness, and response utility during human assessments. It is equipped with advanced RAG capabilities that include verifiable citations, which enhance its utility for enterprise information retrieval applications. Furthermore, the model includes high-level security features designed to protect sensitive business information.
Noteworthy features of Command A include:
- Operational efficiency on two GPUs, significantly lowering computational costs.
- 111 billion parameters optimized for extensive text processing demands in enterprise applications.
- Support for a 256K context length, facilitating effective processing of long-form documents.
- Proficiency in 23 languages, ensuring high accuracy across global markets.
- Exceptional execution in SQL, agentic tasks, and tool-based applications.
- Private deployments being up to 50% more economical than traditional API alternatives.
- Enterprise-grade security to safely manage sensitive data.
The introduction of Command A marks a significant advancement for businesses seeking cost-effective, efficient AI solutions that maintain robust performance standards.
Featured image credit: Kerem Gülen/Midjourney