Cohere made a big move on May 20, open-sourcing its flagship Command A+ model under the Apache 2.0 license, with weights fully available on Hugging Face. The model uses a mixture-of-experts (MoE) architecture with 218 billion total parameters, activating only 25 billion per inference. This is Cohere's first flagship model to be fully open-sourced under Apache 2.0.
Cohere has positioned itself as an enterprise AI provider, serving clients like Citi, PwC, and Oracle. These customers demand the ability to run, modify, and control their own data, making open-source the ultimate solution. By releasing Command A+ under Apache 2.0, Cohere is doubling down on the concept of sovereign AI.
Hardware requirements cut to two H100s
Command A+ is an MoE model with 218B total parameters but only 25B activated per inference, meaning inference costs are based on the active parameters. Cohere's deployment thresholds:
- One B200 GPU is sufficient
- Or two H100 GPUs (with W4A4 quantization)
- Three quantization levels (16-bit BF16, 8-bit FP8, 4-bit W4A4) are all open-sourced on Hugging Face
Two H100s is a significant reduction from the typical 4-8 H100s required for mainstream large models. Cohere claims near-lossless quantization, making this a key threshold for enterprise IT procurement.
Benchmark improvements over previous version
| Benchmark | Command A Reasoning (old) | Command A+ (new) |
|---|---|---|
| τ²-Bench Telecom | 37% | 85% |
| Terminal-Bench Hard Code | 3% | 25% |
| MMMU Pro | - | 63% |
| MathVista | - | 80.6% |
| Artificial Analysis Intelligence Index | - | 37 |
The jump from 37% to 85% on τ²-Bench Telecom and from 3% to 25% on Terminal-Bench Code indicates a major architectural overhaul for agentic tasks, not just an incremental update. The Artificial Analysis score of 37 is comparable to leading open-source models like Llama 4 and Qwen 3.5, but the ability to run on two H100s is rare at this performance level. Output tokens per second are 63% faster than the old version, and first-token latency is reduced by 17%, both critical for enterprise agent scenarios.
What is sovereign AI?
Cohere's blog states that Command A+ is built for "sovereign critical infrastructure." The term sovereign AI means that model weights and data can run entirely within an organization's own data center, without relying on external APIs, local network egress, or third-party data exposure. Customers that care about this include central banks, defense departments, energy/grid companies, large insurers/pension funds, and European enterprises under strict GDPR regulations. For these customers, the biggest barrier to adopting AI is data leaving their premises. No matter how cheap OpenAI or Anthropic APIs are, if an API call must cross the firewall, the deal is off. Cohere is betting on this segment. Fujitsu executive Vivek Mahajan endorsed the model in Cohere's announcement: "Command A+'s mixture-of-experts architecture and strong agentic performance align well with our commitment to deliver innovative, sovereign AI solutions." Fujitsu is a key IT supplier to the Japanese government, serving ministries, financial institutions, and critical infrastructure, and its support reflects Japan's recent push for AI localization.
Native citations: a key detail for enterprises
Command A+ includes a feature called native citations. When the model retrieves information from external tools or databases, it automatically includes grounding spans that link each statement directly to the specific document or database row it references. This is crucial for enterprise knowledge base Q&A, where verifying the source of an answer is a major pain point. Previously, this required external RAG systems to manually bind citations. Cohere has made audit trails a native capability, which is a direct selling point for legal, compliance, and audit departments.
Competition and market positioning
Command A+ is not competing with ChatGPT or Claude for consumer market share; Cohere doesn't fight that battle. Instead, it targets the enterprise open-source market dominated by Meta's Llama. While Llama 4 is open-source, it has commercial use restrictions that require separate licensing for enterprises above a certain scale. Command A+ under Apache 2.0 allows any scale, any scenario, and any commercial use without restrictions. This distinction is significant for enterprise IT decision-makers. Whether this strategy succeeds will depend on Cohere's next ARR report. But as of May 20, 2026, the open-source AI table has a serious player betting on sovereign AI.
Sources: CocoLoop, Introducing Command A+ (Cohere official blog), Cohere cracks lossless quantization and native citations with first full Apache 2.0 licensed open model Command A+ (VentureBeat), Cohere Releases Command A+: An Open-Source Enterprise AI Model Built for Sovereign Critical Infrastructure (Business Wire / Yahoo Finance)