Cohere Command A+: First Apache 2.0 Open Model with Lossless 4-Bit Quantization

Cohere Just Open-Sourced a Frontier Model — What's Command A+?

Cohere, the Canadian AI lab co-founded by "Attention Is All You Need" co-author Aidan Gomez, has released Command A+ — a 218-billion-parameter language model fully licensed under Apache 2.0. This is the company's first fully open-source model release, and it's a major move for enterprise AI builders.

Command A+ is engineered for complex reasoning, multimodal document processing, and agentic workflows. It uses a Sparse Mixture-of-Experts architecture where only 25 billion parameters are active during any generation step, making it dramatically more efficient than dense models of similar capability.

Why Is Apache 2.0 Licensing Such a Big Deal?

Most "open" AI models come with restrictive licenses that limit commercial use or require sharing derivatives. Apache 2.0 is one of the most permissive licenses available — enterprises can use, modify, and deploy Command A+ commercially with minimal restrictions.

This is Cohere's bet on "sovereign AI" — the idea that enterprises, governments, and developers should control their own AI infrastructure without depending on proprietary API providers. You can run it on your own hardware, in your own data center, behind your own firewall.

How Does 4-Bit Quantization Work Without Losing Quality?

The technical breakthrough here is W4A4 quantization — compressing the model to 4-bit precision while maintaining near-lossless performance. Cohere achieved this by only quantizing the MoE experts while keeping attention pathways at full precision, combined with Quantization-Aware Distillation.

The result? Command A+ runs on a single NVIDIA Blackwell B200 GPU or just two H100 GPUs. At low concurrency, it achieves 375 tokens per second with 113ms time-to-first-token — a 63% speed increase and 17% latency reduction over the previous Command A model.

Who Should Pay Attention to This?

If you're an enterprise building AI applications and currently paying per-token to API providers, Command A+ represents a path to owning your AI infrastructure. The combination of open licensing, efficient architecture, and frontier-level performance is unprecedented.

Developers working on agentic systems, document processing, or enterprise search should particularly take note — these are the workloads Command A+ was specifically optimized for.

The Bigger Picture: Cohere and Aleph Alpha Merger

This release comes shortly after Cohere announced a merger with German AI startup Aleph Alpha, signaling a consolidation of non-US AI companies competing against OpenAI, Anthropic, and Google. The combined entity is positioning itself as the sovereign AI provider for European and global enterprises.

FAQ

Q: Can I fine-tune Command A+ for my specific use case? A: Yes, under the Apache 2.0 license you can fine-tune, modify, and deploy Command A+ commercially without restrictions.

Q: How much does it cost to run Command A+ on my own hardware? A: With 4-bit quantization, you need just 2 H100 GPUs — significantly less infrastructure than comparable proprietary models.

Q: How does Command A+ compare to GPT-5.5 or Claude Opus 4.7? A: While proprietary models may have higher raw parameter counts (estimated in the trillions), Command A+ achieves competitive performance through its efficient sparse architecture at a fraction of the deployment cost.

Stay ahead of the AI curve. Follow @AiForSuccess for daily insights.