Cohere 发布 Command A Reasoning 企业级推理模型

Today, we’re introducing Command A Reasoning, our most advanced model for enterprise reasoning tasks. From agentic workflows to end-to-end systems, it outperforms leading privately deployable models in its class, including gpt-oss-120b, DeepSeek-R1 0528, and Mistral Magistral Medium.

Command A Reasoning is purpose-built for enterprise needs, offering highly secure, efficient, and scalable deployment options. For low footprint deployments, it can run on a single H100 or A100 with a context length of 128k. For latency optimized deployments on two or more GPUs the context length scales to 256k. This configurability ensures organizations can make the most of the hardware available to them. The model's long context length makes it ideal for document-heavy workflows and complex multi-step agentic use cases.

Customers can set a token budget to directly manage compute usage and control costs. This eliminates the need to maintain separate reasoning and non-reasoning models, since Command A Reasoning can be used for tasks which require maximum accuracy or configured for greater speed and throughput when efficiency is the main priority.

Command A Reasoning is the core generative model powering North, our secure agentic AI platform. This empowers organizations to deploy custom AI agents and automations on-premises, backed by our most capable reasoning model.

Strong performance across enterprise reasoning tasks

Agentic and multilingual reasoning benchmarks

Command A Reasoning delivers leading results across key agentic benchmarks. It also performs strongly across a range of important business languages, enabling global enterprises to leverage agents with consistent quality.

*Performance evaluated on 3 Agentic benchmarks (higher is better).* *BFCL-v3* *measures a wide range of general single and multihop tool-use agent capabilities.* *Tau-bench* measures performance on two enterprise-relevant customer service agent use-cases, in retail and airline domains, and we measure multilingual capabilities with a carefully constructed in-house translation of tau-bench into 5 languages. Command-a-reasoning outperforms R1-0528, gpt-oss-120b and magistral-medium in all cases.^[1]

Exceptional deep research agent

Command A Reasoning excels at powering end-to-end systems involving chained and hierarchical agents and leveraging the most relevant tools to accomplish tasks. A great example is our Deep Research system, which outperforms similar capabilities from all other leading AI labs.

*Score on the* *DeepResearch Bench* RACE evaluation for our deep research agent on the English question set (higher is better). This measures the overall quality of a long research report generated by a deep research system in terms of instruction following, readability, insight and comprehensiveness. Our system performs the best overall, highlighting its ability to generate strong research reports.^[2]

Deep Research is designed to tackle the complex, in-depth questions that demand more than a quick search. It delivers detailed, well-sourced reports in minutes, which would typically take an employee hours. Our system uses a multi-agent architecture, powered by Command A Reasoning that breaks down the user request into smaller research topics. Then, multiple AI agents work in parallel, searching and analyzing information from a wide array of sources. Finally, the system consolidates the verified findings into a single, well-structured report that directly addresses the original user request. Deep Research is coming soon to North.

Powering North, Cohere’s flagship enterprise AI platform

North enables enterprises that prioritize data security to deploy AI agents and automations at scale within their own infrastructure. On human evaluation scores for a representative set of daily tasks at work, Command A Reasoning consistently outperforms Command A. For customers, these gains translate into more reliable agents, reduced manual intervention, and delivery of highly accurate, actionable results.

The plot shows the average answer satisfaction scores for North when powered by command-a and our new command-a-reasoning model on a wide range of common, but complex, enterprise productivity tasks (higher is better). For these tasks, North needs to successfully understand and navigate multiple sources of Cohere’s real enterprise data (Slack, Google Drive, company emails and web search). The satisfaction score reflects both the accuracy and response quality of the agent. Command-a-reasoning’s enhanced reasoning allows it to better understand a user’s intentions and carry out complex multistep tasks in the enterprise environment, outperforming command-a in all these categories, usually by a wide margin.^[3]

Control and efficiency required for enterprise scale

Command A Reasoning unlocks business applications across industries, combining strong performance, high accuracy, and scalable efficiency. Its low hardware requirements make it practical for private deployments on a single H100 or A100.

The model optimizes compute usage and costs through a user-controlled token budget, enabling seamless adjustment between mission-critical precision and high-throughput tasks. This eliminates the need for separate reasoning and non-reasoning models, allowing enterprises to maximize GPU efficiency and dynamically allocate resources.

In North’s internal evaluations, adjusting the reasoning budget shows a smooth progression in performance from efficient responses at zero reasoning to more in-depth responses at higher reasoning levels. Even with zero reasoning enabled, Command A Reasoning outperforms Command A.

North agent performance on a set of difficult questions requiring multi-hop search (higher is better). Our previous model, Command-a, is shown in pink. Command-a-reasoning has been trained to excel in North, and even with reasoning turned off (the ‘instant’ bar), comfortably outperforms command-a. However, if we allow command-a-reason to think, we see that it improves even more. With more and more thinking time, command-a-reasoning improves further and further. This allows you to fine-grained control over the right amount of latency vs performance for your organization while still only deploying one model.^[4]

Safe in high-stakes environments

Safety is foundational to how we train and evaluate all our models, including Command A Reasoning. This means striking the best balance between ensuring the model doesn’t over-refuse valid requests, and preventing the purposeful propagation of harmful and malicious content online. We focus on five key areas: Child sexual exploitation and abuse (CSEA), self-harm, violence and hate, sexual content, and conspiracy theories.

Cohere internal evaluations show Command A Reasoning strikes the strongest balance among competitors between safety and usefulness. The higher the better for Absolute Safety. The lower the better for Over-Refusal.

Availability

Command A Reasoning is available today on the Cohere platform and for research use on Hugging Face. If you are interested in private or on-prem deployments, please contact our sales team for bespoke pricing.

“We’re excited to partner with Cohere to integrate their Command A Reasoning model into our generative AI hub in SAP Business Technology Platform. This powerful model will enhance SAP’s generative AI capabilities, empowering customers, partners, and developers to build innovative, secure agentic applications tailored to their unique environments. Together, we’re unlocking new possibilities for enterprise reasoning tasks.” — Dr. Walter Sun, SVP and Global Head of AI, SAP SE

^[1] We run all models with their highest available reasoning setting.
On BFCL, We evaluate command-a-reasoning using the Function Calling (FC) setting. For competitors, we report their score on the official BFCL leaderboard if available, or otherwise benchmark them using the official BFCL codebase, reporting the highest of their prompted and FC evaluation settings.
On Taubench, we report the average of the pass^1 score over 10 runs for command-a-reasoning. For competitors, where available, we take the officially reported numbers (R1-0528, GPT-oss) otherwise we run the tool-use api against the official tau-bench implementation, reporting the average pass^1 over 10 runs. In all cases, we report the average of airline and retail.
On M-taubench, we run all models against and report the average score across Ja, Ko, Ar, Es, Fr, and En for both retail and airline, across 10 runs.
^[2] Our deep research agent is a hierarchical web-search agent which breaks a problem down into multiple subproblems, and recursively researches each with a sub-agent. When all subagents have finished, we take their subreports and generate a final report with a number of steps of iterative refinement, making sure our final report is strong. We use an internally developed agent and web search, leveraging our expertise in retrieval, reranking, retrieval-augmented generation and agents. We report the RACE scores for competitor models on the english questions by rerunning the official RACE evaluation on their english deep research reports on the DeepResearchBench leaderboard website.
^[3] Annotators were asked to rate their satisfaction score based on a comprehensive internally-developed rubric, summarizing aspects of response quality, appropriateness, correctness and latency. The total evaluation set is composed of 112 questions, and each answer is annotated by 6 annotators before we aggregate their scores.
^[4] This evaluation uses an automatic correctness score utilizing LlamaIndex, with respect to known correct reference answers.