Overview
Command A Translate (model ID command-a-translate-08-2025) is Cohere's dedicated machine-translation model, announced on August 28, 2025. It is a 111-billion-parameter dense decoder-only Transformer built on top of Cohere's flagship Command A (03-2025) model and specialized for high-quality translation across 23 business languages: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Chinese, Arabic, Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian.
The model has a 16K-token context window (8K for input, 8K for output) and is text-only. Cohere positions Command A Translate as an enterprise-grade, secure alternative to general-purpose chat models and standalone translation services, emphasizing data control: it can run through the Cohere platform's API, as a private deployment, or fully on-premises on one or two A100/H100 GPUs. The open weights were released for research on Hugging Face under a CC-BY-NC license plus Cohere's Acceptable Use Policy, with commercial use handled through a Cohere agreement.
A distinguishing feature is Deep Translation, an agentic, multi-step refinement workflow that iteratively improves draft translations before delivery for more complex translation tasks. Cohere reports that on the WMT24++ benchmark (English into 55 languages and dialects, scored with the xCOMET-XL quality metric) Command A Translate outperforms the systems it was compared against — including GPT-5, DeepSeek-V3, DeepL, and Google Translate — on all evaluated languages except Hebrew and Hindi, and that adding Deep Translation pushes quality higher still.
| Released | 2025-08-28 |
|---|---|
| License | CC-BY-NC 4.0 (non-commercial) plus Cohere's Acceptable Use Policy for the open weights; commercial use requires a Cohere agreement |
| Weights | Open weights |
| Parameters | 111B |
| Context | 16K tokens (8K input + 8K output) |
| Max output | 8K tokens |
| Architecture | Dense decoder-only Transformer, built on Command A (c4ai-command-a-03-2025). Uses interleaved attention: three layers of sliding-window attention (window 4096) with RoPE, then a fourth layer of global attention without positional embeddings. Specialized for translation and tuned with Direct Preference Optimization (DPO) over a difficulty-filtered training set. |
| Knowledge cutoff | Not published by Cohere |
| Modalities | text |
| Status | Available. Released August 28, 2025. Open-weight research release on Hugging Face; commercial use via the Cohere platform or private/on-prem deployment. |
Strengths
- High translation quality across 23 enterprise languages, reported by Cohere to beat GPT-5, DeepSeek-V3, DeepL, and Google Translate on most WMT24++ language pairs
- Deep Translation: an agentic, multi-step refinement workflow that raises quality on harder content
- Data control and security: available as a private deployment or on-prem install (one or two A100/H100 GPUs)
- Open weights released for research on Hugging Face
- Specialized model rather than a general chat LLM, tuned specifically for translation with difficulty-filtered data and DPO
Best for
- Enterprise document and content translation across 23 languages
- Secure or regulated workflows that require private/on-prem deployment and full data control
- Localization pipelines where higher quality justifies an agentic Deep Translation refinement pass
- Translating business-critical content (legal, financial, technical) in supported non-Latin scripts such as Japanese, Korean, Chinese, Arabic, and Hebrew
- Research and evaluation of state-of-the-art machine translation using the open weights
How to access
| Provider | Model ID |
|---|---|
| Cohere ↗ | command-a-translate-08-2025 |
FAQ
What is Command A Translate?
Command A Translate (model ID command-a-translate-08-2025) is Cohere's dedicated machine-translation model, released August 28, 2025. It is a 111B-parameter decoder-only Transformer built on Command A (03-2025) and specialized for high-quality translation across 23 languages.
How many languages does it support?
23 languages: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Chinese, Arabic, Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian.
What is Deep Translation?
Deep Translation is an agentic, multi-step refinement workflow that iteratively improves draft translations before delivery. Cohere reports it raises quality further on complex translation tasks, adding roughly +2 xCOMET-XL over the best competing system it was measured against.
Is Command A Translate open weights, and how much does it cost?
The weights are released for research on Hugging Face under a CC-BY-NC license plus Cohere's Acceptable Use Policy; commercial use requires a Cohere agreement. Cohere does not publish per-token API pricing for the Translate model — it is offered through the Cohere platform or as a private/on-prem deployment via Cohere's sales team.