Command A+: An Open Enterprise MoE on Two H100s

Cohere released Command A+ on May 20, 2026 as an Apache-2.0, privately deployable enterprise model. The technical launch article describes a sparse mixture of experts with 218 billion total parameters and 25 billion active, 128K input context, 64K maximum generation, image input, reasoning, tool use, and support for 48 languages. Quantized deployment can fit on two H100s or one B200.

This is not simply a bigger Command A. Cohere consolidated capabilities that previously lived in separate reasoning, vision, and translation models. It also targeted a specific operational constraint: regulated organizations need an agentic model they can inspect, run in a private environment, and serve without a vast accelerator cluster. The benchmark evidence shows large gains over Command A Reasoning, but the more durable contribution may be the deployability-to-capability ratio.

Cohere's parallel sovereign-infrastructure announcement makes that deployment goal explicit. As with every vendor release, the claims still need a task-level pilot using the controls in our frontier model evaluation guide.

What is actually open

Command A+ weights are available under Apache 2.0 in several precisions. Cohere documents BF16, FP8, and W4A4 variants and supports common inference frameworks. That is a meaningful open release: organizations can download the checkpoint, inspect its files, serve it on their infrastructure, and modify it under a permissive license.

Open weights do not disclose every training example or guarantee freedom from policy obligations. Model governance still requires provenance records, safety testing, license review, and a documented deployment configuration. The hosted Cohere service and self-managed weights are different products even when they share a model name.

The parameter description also needs precision. Cohere's narrative article says 24 billion active in one place, while the detailed snapshot and model documentation say 25 billion. We use 25B active because it is the value in the structured specification. The small discrepancy is exactly why model records should preserve source and update date.

Architecture and serving envelope

The 218B/25B sparse design aims to provide large stored capacity while keeping token compute closer to a much smaller dense model. The 128K input window is ample for enterprise retrieval and document packets, and the 64K output limit supports long analytical or agent traces. Image input allows the same model to interpret charts, PDFs, slides, and screenshots.

Minimum hardware claims assume aggressive quantization. Two H100s or one B200 at W4A4 is an attractive floor, but a production service also needs KV-cache memory, batching headroom, redundancy, and acceptable concurrency. Quantized quality must be checked on the organization's own workloads, especially multilingual text, tables, and numeric reasoning.

Cohere reports up to 63 percent higher output throughput and up to 17 percent lower time to first token than Command A Reasoning at matched quantization and concurrency. W4A4 adds a further speed increase in the published setup. Those are controlled vendor measurements on an HGX B200 node, not promises for every engine.

The benchmark gains

Command A+ combines public benchmark results and internal North evaluations:

The headline τ² gain is large enough to warrant attention. It also depends on simulated users and the benchmark's telecom environment. Terminal-Bench Hard used the Terminus-2 methodology described by Cohere. Multimodal results follow standard datasets, but image preprocessing and output constraints still matter.

Enterprise performance is not one public leaderboard

Cohere also reports internal North gains: agentic question answering improved 20 percent, spreadsheet analysis quality 32 percent, and memory quality reached 54 percent versus 39 percent for Command A Reasoning. These tasks reflect the workflows the model was designed to power, which makes them practically relevant and less independently reproducible.

Internal benchmarks are not inherently invalid. They can sample proprietary workloads that public tests miss. They become persuasive when evaluation cases, graders, and failure categories are disclosed. Buyers should ask Cohere for a task-level methodology and then reproduce a representative subset with their own files and permissions.

In document agents, correctness must be field-level. A fluent summary that changes a number, misses a footnote, or attributes a statement to the wrong source is a failure. In spreadsheets, verify formulas and references. In retrieval, score citation support rather than topical similarity.

Multilingual support needs native acceptance tests

Command A+ expands the Command line from 23 to 48 languages, including Persian, Arabic, Japanese, Korean, Hindi, and all official EU languages. Cohere reports tokenizer compression improvements of 20 percent for Arabic, 16 percent for Korean, and 18 percent for Japanese. Fewer tokens can reduce cost and preserve more content inside a fixed context.

Tokenizer efficiency is not language quality. Persian deployment should test formal and conversational registers, half-space behavior, Arabic-versus-Persian characters, calendar and number conventions, mixed-script identifiers, and right-to-left document order. Translation evaluation should preserve names, amounts, citations, and formatting.

WMT-style automatic scores are helpful for broad comparison, but regulated communications require native reviewers and domain glossaries. A sovereign model is valuable only if the local language experience is accurate and culturally appropriate.

Why the hardware claim matters

Many open-weight flagships require eight or more high-memory accelerators before KV cache and redundancy. Command A+'s claim of practical inference on one B200 or two H100s lowers the threshold for a private pilot. It may let a regulated team keep documents inside its network without accepting the latency and privacy trade-offs of an external API.

The serving design still needs engineering. Choose an inference engine, pin kernels, test quantization, set context limits, and monitor routing or numerical instability. Capacity planning should use target concurrency and prompt distributions, not the minimum configuration in a model card.

Speculative decoding adds another reported 1.5–1.6x speedup. Its benefit depends on draft acceptance and batch size. Measure it with the same prompts and accepted quality; do not multiply every vendor maximum into a theoretical speedup that no workload achieves.

Best-fit workloads and weak fits

Command A+ is strongest on private enterprise agents, retrieval-augmented generation, multilingual document analysis, SQL and spreadsheet work, and tool-rich processes. The unified model can simplify routing among separate Command checkpoints and reduce operational overhead.

It is a weaker fit when a tiny edge model is required, when 128K context is insufficient, or when the organization cannot operate a 218B sparse checkpoint. A hosted frontier model may be cheaper for low volume. A specialized OCR system may outperform general vision on scanned documents. Open weights create options; they do not make self-hosting automatically economical.

Compare total cost of ownership: accelerators, orchestration, observability, security, upgrades, and human review. The open-model operations guide lays out the release and rollback controls needed after download.

How to pilot Command A+

Build a bilingual, multimodal test set from authorized examples. Run BF16 or FP8 as the reference, then compare W4A4 at equal prompts. Record exact-match fields, supported citations, tool success, latency percentiles, token counts, and accelerator utilization. Include adversarial documents, prompt injection, conflicting sources, and tasks that require refusal or escalation.

For agents, isolate tools in a sandbox and require confirmation for material actions. For RAG, test retrieval separately from generation so model improvements are not confused with index changes. For translation, use native human evaluation. For code and SQL, execute the output against disposable systems.

Command A+ earns its place in the series because the release joins strong benchmark progress with a concrete private-serving story. The next decision must be based on a controlled local trial, not the appeal of “sovereign AI” alone.

Source notes — reviewed 2026

• Cohere's technical Command A+ release supplies specifications, benchmark values, throughput tests, and evaluation footnotes.
• Cohere's sovereign-infrastructure announcement explains the enterprise and licensing position.
• Cohere's model documentation records the structured model properties and release date.
• Cohere's API changelog identifies the hosted model ID and availability.
• The official Hugging Face organization is the first-party distribution channel for Cohere's downloadable checkpoints.