Are you looking for smarter insights delivered directly to your inbox? Sign up for our weekly newsletters and receive essential updates tailored for leaders in enterprise AI, data, and security. Subscribe now!
Cohere’s New Model Launch
Today, I caught up on some exciting news amidst a busy meeting schedule: Cohere, the Canadian startup co-founded by Aidan Gomez, a co-author of the influential Transformer paper, has launched its first reasoning large language model (LLM), named Command A Reasoning. This model is designed to make generative AI products work seamlessly, powerfully, and securely for enterprises. Early benchmarks, technical specifications, and tests indicate that this model excels in flexibility, efficiency, and reasoning capabilities.
Capabilities and Features
Cohere claims that Command A Reasoning can automate various tasks at scale within secure enterprise environments, including customer service, market research, scheduling, and data analysis. Although it is a text-only model, it can easily integrate with multimodal models and tools, which is one of its primary advantages.
Challenges in AI Scaling
Enterprise AI is facing challenges such as power limitations, increasing token costs, and inference delays. Join our exclusive salon to learn how leading teams are transforming these challenges into strategic advantages by:
– Turning energy into a competitive edge
– Architecting efficient inference for significant throughput gains
– Unlocking sustainable AI systems for better ROI
Secure your spot to stay ahead: https://bit.ly/4mwGngO
Access and Pricing
While the model is available for researchers to use for non-commercial purposes, enterprises must pay for access, and Cohere does not publicly disclose its pricing due to customized deployment options. Following its recent funding round of $500 million, Cohere was valued at $6.8 billion. Command A Reasoning is specifically tuned for enterprises with extensive document libraries, lengthy email chains, and workflows that cannot tolerate inaccuracies.
Technical Specifications
The model supports up to 256,000 tokens on multi-GPU setups, which is a considerable size, comparable to OpenAI’s GPT-5. Command A Reasoning features 111 billion parameters and is trained with a focus on tool use and multilingual performance, supporting 23 languages, including English, French, Spanish, Japanese, Arabic, and Hindi. This multilingual capability is crucial for global enterprises that require consistent quality across different markets.
Integration with North Platform
Command A Reasoning integrates seamlessly into North, Cohere’s new platform for deploying AI agents and automations on-premises. This allows enterprises to create custom agents that operate entirely within their infrastructure, maintaining control over data flows while leveraging advanced reasoning capabilities. Cohere has strategically identified common functions across enterprises—such as onboarding, market research, and analysis—and designed its model to support these workflows automatically.
Token Budget Feature
Similar to other recent reasoning models, including Nvidia’s Nemotron-Nano-9B-v2, Command A Reasoning introduces a token budget feature that enables users or developers to specify the amount of reasoning allocated to specific inputs and tasks. A lower budget results in faster, cheaper responses, while a higher budget allows for deeper and more accurate reasoning. The Hugging Face release even allows users to toggle reasoning on or off through a simple parameter.
Developers can utilize the model in “reasoning mode” for maximum performance or switch it off for lower latency tasks without needing to change models.
Performance Benchmarks
How does Command A Reasoning perform in real-world applications? Cohere’s benchmarks reveal a clear advantage. In enterprise reasoning tasks, Command A Reasoning consistently outperforms competitors like DeepSeek-R1 0528, gpt-oss-120b, and Mistral Magistral Medium. It also excels in multilingual benchmarks, which is essential for global businesses.
The token budget system proves effective, as satisfaction scores improved steadily with increased budgets. Even with minimal reasoning, Command A Reasoning surpassed its predecessor, and at higher budgets, it continued to pull ahead. In deep research evaluations, such as the DeepResearch Bench—which assesses instruction following, readability, insight, and comprehensiveness—Cohere’s model outperformed offerings from Gemini, OpenAI, Anthropic, Perplexity, and xAI’s Grok.
Practical Applications
Beyond impressive benchmarks, Command A Reasoning is designed for practical use. Cohere specifically trained it for conversational tool use, allowing it to call APIs, connect to databases, or query external systems during tasks. Developers can define tools using JSON schema and incorporate them into chat templates in Transformers, facilitating smoother integration into existing enterprise systems.