Command A+ Guide: Cohere Goes Sparse, Apache 2.0 and 48 Languages

Cohere's earlier open releases came with a catch: the weights were downloadable, but the licence was non-commercial. Command A+ removes the catch. It ships under Apache 2.0, which makes it the first Command model a company can download, fine-tune and run in a product without a separate agreement.

That licence change, more than any benchmark, is what makes Command A+ worth a look. It is a mid-tier model by 2026 standards, not a frontier one. Its pitch is a capable, permissively licensed, multilingual model that fits on hardware an enterprise can actually own.

What changed from Command A

Cohere announced Command A+ on 20 May 2026, with weights published as command-a-plus-05-2026; OpenRouter's listing followed in September. Cohere says it unifies the capabilities of the earlier Command A variants in one model, so vision input, reasoning, translation and agentic tool use no longer need separate checkpoints.

The architecture is the bigger break. Every earlier Command model was dense. Command A+ is a sparse mixture-of-experts model with 218B total parameters and 25B active per token: 128 experts, eight routed per token plus one shared expert. The practical consequence is that its compute per token is far below that of the dense 111B Command A, while its total capacity is roughly double. Active vs total parameters covers why the active figure is the one that sets serving cost.

Cohere claims up to 63 per cent higher output tokens per second and up to 17 per cent lower time to first token than Command A Reasoning, its previous reasoning model. Those are vendor figures without published test conditions, but the direction is what a move from dense to sparse would predict.

The hardware story is the real pitch

Cohere says Command A+ deploys on as few as one B200 or two H100 GPUs. Read the fine print: that figure is for the W4A4 release on Hugging Face, meaning weights and activations both quantised to 4 bits. At BF16, 218B parameters take roughly 436 GB, and the model card puts the minimum at four B200s or eight H100s; FP8 halves that to two B200s or four H100s.

4-bit activations are a more aggressive choice than the 4-bit weight-only quantisation most self-hosters are used to. The model card says all three precisions show negligible differences in benchmark quality, but your tasks are not its benchmarks. If you plan to self-host on two H100s, evaluate on the W4A4 build itself rather than on Cohere's hosted API, which may not run the same precision.

The combination of Apache 2.0 and a two-card footprint is aimed squarely at organisations that need data to stay on their own hardware: public sector, defence, healthcare, regulated finance. Open-weight model licences compared puts Apache 2.0 alongside the MIT and custom licences elsewhere in the field.

The agentic jump, with caveats

Cohere's launch post reports a large step on agent benchmarks, from a low base:

• τ²-Bench Telecom (customer-service agent tasks) — 85 per cent, up from 37 for Command A Reasoning
• Terminal-Bench Hard (command-line coding tasks) — 25 per cent, up from 3

The first is a strong result on a benchmark that resembles real enterprise agent work: following a policy, calling tools, and resolving a ticket. The second is a big relative gain that still leaves the model well behind 2026's specialist coding models. Do not pick Command A+ for autonomous coding.

Independent aggregate rankings agree. At the time of writing, Artificial Analysis places it below the median of comparable models on its Intelligence Index, while measuring its output speed well above the median. Fast and moderately capable is a fair summary of where it belongs.

48 languages is not a footnote

Command A+ supports 48 languages, more than double the 23 that Command A covered, and the model card's list includes every official EU language. The list includes Irish, Maltese, Estonian, Icelandic, Filipino, Punjabi, Tamil and Telugu, which many models handle only incidentally.

For a European public-sector deployment, or any product that must serve smaller EU languages at the same quality as English, this is the most differentiated thing about the model. Test it on your own language pairs with domain terminology, because broad coverage does not guarantee equal quality across all 48. LLM translation covers how to test placeholders, glossaries and register.

Price and the context arithmetic

On OpenRouter at the time of writing, Command A+ lists at $0.30 per million input tokens and $1.50 per million output, with cached input at $0.15. Against Command A's listed $2.50 and $10, that is more than eight times cheaper on input and over six times cheaper on output, for a newer model with more capabilities.

The context figure needs care. Cohere specifies 128K tokens of input and 64K of output. Some catalogues add these and list a single 192K window, which is misleading if you plan to send 190K tokens of documents. The input ceiling is 128K, half of Command A's 256K.

The reasoning mode writes its thinking between tags before the answer, and those tokens bill as output. Leave it off for retrieval answers and extraction, and turn it on for multi-step agent planning where you have measured that it helps.

When to pick it, and when not to

• Pick Command A+ for enterprise RAG, policy-following agents, document processing in many languages, and any case where you need a commercially usable model on your own hardware.
• Pick Command A only when you need more than 128K tokens of input in one call, or have a validated pipeline you cannot yet re-test.
• Pick something else for autonomous coding or frontier-level reasoning, where Command A+ is simply not competitive.

The fastest way to decide is to run your own retrieval and agent tasks against it at both reasoning settings, and on the W4A4 weights if self-hosting is the plan.