Cohere cracks lossless quantization and native citations with first full Apache 2.0 licensed open model Command A+

Cohere unveiled Command A+, a 218-billion-parameter language model released under the Apache 2.0 open license. The model represents the company's first fully open-source offering and arrives weeks after Cohere announced its merger with German AI startup Aleph Alpha.

Command A+ targets enterprise use cases requiring complex reasoning, multimodal document processing, and agentic workflows. The model introduces two technical innovations. First, lossless quantization allows the model to run efficiently on consumer hardware without degrading performance. Second, native citation capabilities let the model attribute claims directly to source documents, addressing a core limitation in current language models that often generate unsourced statements.

The lossless quantization breakthrough matters because it lets businesses deploy a 218-billion-parameter model on standard infrastructure, reducing the computational barrier for organizations without dedicated AI hardware clusters. This contrasts sharply with most frontier models that require specialized GPU infrastructure to run.

Native citations attack the hallucination problem differently than retrieval-augmented generation. Rather than fetching documents externally, Command A+ embeds citation tracking into its reasoning process. When the model makes a claim, it points to the specific source, enabling users to verify outputs directly.

Cohere co-founder Aidan Gomez, a co-author of the "Attention Is All You Need" paper that spawned modern transformers, positioned the release as democratizing powerful reasoning capabilities. The Apache 2.0 license removes usage restrictions, letting developers build commercial products without licensing fees.

The timing signals Cohere's strategic shift toward open-source competition with Meta's Llama models. While OpenAI and Anthropic maintain closed API models, Cohere and Meta have chosen openness as a competitive moat, betting that developers will build more extensively with transparent models they control.

The merger with Aleph Alpha, combined with this