Cohere Unveils Command A+: Open-Source 218B MoE Model with Lossless Quantization and Native Citations

A New Chapter in Enterprise AI

Canadian AI lab Cohere, co-founded by former Googler and “Attention Is All You Need” co-author Aidan Gomez, has a history of pushing boundaries. After announcing a merger with German startup Aleph Alpha, the company now introduces Command A+—a 218-billion-parameter language model designed for complex reasoning, multimodal document processing, and agentic workflows. What sets this release apart is not just its technical prowess but its unprecedented openness: for the first time, Cohere offers a model under the permissive Apache 2.0 license, making the weights freely available on Hugging Face. This move underscores a bet on “sovereign AI”—the idea that enterprises, governments, and developers can run frontier-level AI in secure, controlled environments without sacrificing performance.

Open Access Under Apache 2.0: Empowering Sovereign AI

By adopting the Apache 2.0 license, Cohere breaks from its previous proprietary approach. The model weights are hosted on Hugging Face, allowing anyone to download, modify, and deploy Command A+. This aligns with a growing demand for sovereign AI—where organizations retain full control over their data and infrastructure. Aidan Gomez announced the decision via X (formerly Twitter), highlighting that enterprises now have a path to adopt cutting-edge AI without vendor lock-in or data privacy concerns. The move is particularly significant for governments and regulated industries that require on-premises deployment or air-gapped environments.

Sparse Mixture-of-Experts Architecture for Efficiency

Command A+ uses a decoder-only Sparse Mixture-of-Experts (MoE) Transformer. While the model has 218 billion total parameters, only 25 billion are active during any inference step. This design dramatically reduces computational overhead compared to dense models like GPT-5.5 or Claude Opus 4.7, which are estimated to have trillions of parameters. The MoE approach directs each query only to the most relevant “expert” subnetworks, keeping the rest dormant. As a result, the model retains vast knowledge and nuanced reasoning while operating at speeds and energy costs typical of far smaller models.

Active Parameters and Inference Speed

The sparse architecture is key to Command A+’s efficiency. With only about 11% of its parameters active at a time, the model achieves inference speeds rivaling models with far fewer total parameters. This makes it suitable for real-time applications, agentic loops, and high-throughput document processing—all without requiring the massive GPU clusters needed for dense giants.

Advanced Quantization Without Quality Loss

Quantization reduces a model’s memory footprint by lowering the precision of its parameters. Command A+ supports multiple formats: 16-bit (BF16), 8-bit (FP8), and a highly compressed 4-bit (W4A4). The 4-bit quantization is the standout technical achievement. Typically, reasoning models suffer a “quantization tax”—compression leads to degraded performance on complex tasks. Cohere mitigated this through specialized techniques that preserve fidelity even at extreme compression levels.

W4A4: Breaking the Quantization Tax

Command A+’s W4A4 quantization delivers lossless performance in benchmarks, according to Cohere. This means enterprises can run the model on lower-cost hardware (e.g., consumer-grade GPUs or edge devices) without compromising accuracy. The ability to deploy high-performance models on modest infrastructure is a game-changer for cost-sensitive deployments and on-premises use cases.

Built-In Native Citations for Trustworthy Outputs

Command A+ introduces native citations—a mechanism that automatically attributes information sources within generated text. Unlike post-hoc retrieval methods, this is integrated into the model’s architecture. When the model produces a claim, it simultaneously references the specific passages from its training data or provided context that support it. This feature enhances transparency and trust, especially in enterprise applications where auditability is critical.

How Native Citations Work

During generation, Command A+ outputs inline citations that trace back to source documents. In a document analysis task, for example, the model might state a fact and immediately link to the paragraph where that fact appears. This reduces hallucination and makes it easier for users to verify outputs. For regulated industries like legal or healthcare, native citations provide a dependable trail of evidence.

Multimodal Document Processing and Agentic Workflows

Command A+ is optimized for handling diverse document types—text, tables, images, and code—and for powering autonomous agents. Its multimodal capability means it can extract insights from PDFs, spreadsheets, and slides without separate pipelines. In agentic workflows, the model can reason across steps, use tools, and make decisions with minimal human intervention. Combined with its efficiency and open license, Command A+ positions itself as a versatile backbone for next-generation enterprise automation.

Enterprise Use Cases

Potential applications include automated compliance auditing, complex research synthesis, intelligent document review, and customer support agents that handle multifaceted queries. The open license also enables fine-tuning for domain-specific tasks, further expanding its utility. With 25B active parameters and lossless quantization, Command A+ balances power and practicality for deployment at scale.

Looking Ahead: The Open Model Revolution

Cohere’s decision to open-source Command A+ under Apache 2.0 reflects a broader industry shift. By combining a sparse MoE architecture with state-of-the-art quantization and native citations, the company offers a model that is both powerful and accessible. Enterprises no longer have to choose between capability and control. As sovereign AI gains momentum, Command A+ may well become a reference point for what open, enterprise-grade AI can achieve.