Cohere released Command A+ today, an open-source enterprise language model that consolidates the capabilities of its entire Command family into a single, efficient system. The model is available under an Apache 2.0 license, which means anyone can use it commercially without restriction.
The specifications tell an interesting story about where the company thinks the industry is headed. Command A+ uses a Mixture-of-Experts architecture with 218 billion total parameters, but only 25 billion are active for any given computation. That design choice matters because it lets the model fit onto two H100 GPUs while maintaining performance that competes with much larger systems.
Significant Benchmark Gains
Cohere reports substantial improvements over its previous Command A Reasoning model. On τ²-Bench Telecom, a customer service agent benchmark, scores jumped from 37% to 85%. Agentic coding performance on Terminal-Bench Hard increased from 3% to 25%. The company also claims gains across non-agentic reasoning, instruction following, and general code generation.
These numbers suggest the model could handle the kinds of tasks enterprises actually care about: customer service automation, internal tool orchestration, and complex document workflows. Whether the benchmarks translate to real-world performance is always the open question, but the trajectory is encouraging.
Why This Release Matters
Cohere has spent the past year building North, its enterprise workspace for deploying agentic AI. According to the company, that work drove much of the innovation behind Command A+. The goal was to create a unified model that simplifies deployment, runs locally, and synthesizes capabilities from across the Command family.
Nick Frosst, Cohere's co-founder, framed the release in terms of sovereignty. As open-source AI development concentrates in fewer jurisdictions, organizations operating critical systems face growing concerns about transparency, security, and vendor dependence. Command A+ addresses these by letting enterprises deploy the model wherever sensitive data resides, whether in a VPC, on-premises, or fully air-gapped.
Technical Details Worth Noting
The company introduced several optimizations alongside the base model. Command A+ uses a new tokenizer that delivers what Cohere describes as substantial compression improvements. Fewer tokens for the same response means lower inference costs.
Speculative decoding, optimized specifically for the MoE architecture, provides an additional 1.5x to 1.6x inference speedup for both text and multimodal inputs. The model also supports vision capabilities, processing images alongside text for enterprise use cases like document analysis and chart interpretation.
The model weights are available on Hugging Face in several quantization formats. Cohere recommends the W4A4 quantization for most users, citing superior speed and latency characteristics with minimal quality degradation.
The Competitive Landscape
This release positions Cohere as a serious contender in the open-source enterprise AI space. The company has always emphasized efficiency over raw scale, and Command A+ continues that philosophy. Running a frontier-class model on two GPUs instead of eight fundamentally changes the economics of private deployment.
For organizations that need to maintain full control over their AI infrastructure, whether for regulatory compliance, data sensitivity, or strategic reasons, options have been limited. Meta's Llama models are powerful but less enterprise-focused. Mistral offers efficiency but lacks the agentic capabilities that Command A+ provides.
Cohere's approach of building models specifically for enterprise deployment, rather than adapting consumer-focused systems, may prove to be a meaningful differentiation. The company's partnership with Oracle for agentic AI deployment and its focus on sovereign infrastructure suggest a long-term strategy that goes beyond just releasing models.
Command A+ is available now through Cohere's API, Model Vault for managed inference, and via direct download from Hugging Face.
