Cohere launches Command A+ open-source model to challenge Silicon Valley dominance

The Canadian artificial intelligence company Cohere has released its most powerful language model to date, Command A+, under the highly permissive Apache 2.0 open-source license[1]. This release represents a significant shift for the company, serving as its first fully open-source, enterprise-grade model[1][2]. For a long time, leading AI laboratories have restricted access to their most capable systems behind commercial application programming interfaces or semi-restrictive research licenses[3]. By providing the open weights of Command A+ freely on public repositories, Cohere is signaling a major commitment to digital sovereignty and open science[4][3]. The launch comes on the heels of the company’s recent landmark merger with German AI pioneer Aleph Alpha, a deal valued at approximately twenty billion dollars that established a transatlantic alliance specifically designed to challenge the dominance of American tech conglomerates[5][6].

At the technical core of Command A+ is a highly optimized Sparse Mixture-of-Experts Transformer architecture[1][3]. Unlike dense neural networks that activate all parameters for every calculation, this architecture routes incoming queries only to specific sub-networks, or experts, best suited to handle them[3]. While Command A+ houses a massive two hundred and eighteen billion total parameters, it only activates twenty-five billion parameters during any given generation step[7][3]. This design dramatically slashes the compute resources required for inference, allowing the model to deliver frontier-class performance with a fraction of the hardware footprint of its massive, proprietary competitors[1][3]. This efficiency makes the model highly practical for organizations with constrained compute environments, presenting a viable alternative to the resource-heavy systems operated by larger competitors[1][8].

To further lower the barrier to entry, Cohere has engineered Command A+ with native support for advanced quantization techniques[7][3]. In artificial intelligence, quantization compresses model weights to reduce memory usage, though it often comes at the cost of reasoning quality[3][9]. Cohere, however, has achieved near-lossless quantization with its four-bit format, known as W4A4[7][3]. The company has demonstrated that the difference in benchmark quality between the full-precision sixteen-bit version and the highly compressed four-bit version is virtually negligible[7]. This means that an enterprise can run a model of this magnitude on highly accessible hardware[7][8]. For example, the four-bit version can operate efficiently on a single Nvidia B200 graphics processing unit or just two H100 units, radically reducing the infrastructure cost for organizations wanting to deploy AI locally[7][9].

Beyond its hardware efficiency, Command A+ introduces substantial upgrades in speed, reasoning, and functionality[10][9]. It achieves more than double the output speed and a thirty percent reduction in latency compared to its predecessor, Command A[9][11]. Equipped with a one hundred and twenty-eight thousand token context window capable of producing up to sixty-four thousand output tokens, the model is built to tackle complex, multi-step agentic workflows and massive document analysis[7][10]. Crucially, the model is fully multimodal, accepting both text and image inputs[7][12]. This allows it to ingest, parse, and reason over highly complex visual documents, charts, and mathematical figures[13]. In initial benchmarks, Command A+ has shown strong capabilities in multimodal document processing, performing exceptionally well on standard evaluations such as MathVista, MMMU, and CharXiv[13].

Another defining feature of Command A+ is its sophisticated implementation of Retrieval-Augmented Generation and native citations[14][3]. One of the primary obstacles to deploying generative AI in professional environments is the risk of hallucinations, where models confidently invent false information[15]. To address this, Command A+ natively embeds special tags within its outputs, directly linking every factual assertion it makes to a specific source document, database row, or external reference[14]. This native citation mechanism enables developers to construct highly transparent search-and-retrieval pipelines, giving human operators the ability to instantly verify the AI’s conclusions[14]. Coupled with its multilingual support for forty-eight languages, including all official European Union languages, the model offers global businesses a highly precise tool for cross-border operations[10].

The decision to release Command A+ under an Apache 2.0 license is deeply intertwined with the geopolitical concept of sovereign AI[1][4]. In the current technological landscape, a small handful of American and Chinese corporations control the primary pipelines of generative AI, forcing international enterprises and public sector agencies to rent intelligence via cloud-hosted APIs[1][3]. This setup presents severe security and compliance hurdles for governments and regulated sectors, such as healthcare, finance, and defense, which must adhere to strict data localization laws[1][5]. Because Command A+ is fully open-source, organizations can download the model weights and deploy them entirely within their own secure environments, whether that means a Virtual Private Cloud, on-premises data centers, or completely air-gapped facilities[1]. This level of deployment freedom ensures total data sovereignty and operational control, protecting sensitive national and corporate data from foreign jurisdiction[1].

This focus on technological independence is a logical continuation of Cohere’s strategic positioning[1][5]. The recent merger with Aleph Alpha, which was backed by both the German and Canadian governments and funded in part by Schwarz Digits, was explicitly pitched as a means to build a robust sovereign alternative to the Silicon Valley ecosystem[5][6]. The launch of Command A+ provides the actual software foundation to back up this geopolitical alliance[1]. By providing a model that can be hosted locally within Europe or Canada, Cohere is directly catering to European and public sector clients who are wary of vendor lock-in and foreign data snooping[1][5]. Company leadership has emphasized that prioritizing sovereignty allows institutions to truly own their cognitive infrastructure, rather than remaining dependent on a few dominant tech companies to dictate how their internal AI systems function and evolve[1].

Ultimately, the release of Command A+ marks a pivotal moment in the democratization of enterprise AI[1]. By combining a highly sophisticated Sparse Mixture-of-Experts architecture, cutting-edge lossless quantization, and a fully open Apache 2.0 license, Cohere has dismantled the notion that high-performance, secure AI must be restricted to proprietary cloud platforms[14][1][3]. As organizations worldwide navigate the dual pressures of scaling AI capabilities while managing strict compliance and rising infrastructure costs, Command A+ offers a compelling blueprint for the future[1][13]. It demonstrates that the path to widespread AI adoption lies not in centralized, closed-source monopolies, but in giving governments and enterprises the open tools they need to build, control, and secure their own intelligence[1].