Command A
Command A is a 111B-parameter model from Cohere, released on March 13, 2025, for tool use, retrieval-augmented generation (RAG), agents and multilingual use. It has a 256K-token context length and runs on two A100 or H100 GPUs.
Released
March 13, 2025
Type
llm
License
open-weight
CC-BY-NC 4.0 (plus Cohere Labs Acceptable Use Policy)
Context
256,000 tokens
Pricing
Input
$2.50 / 1M tokens
Output
$10.00 / 1M tokens
List price on the developer's own pages, checked . Pay-as-you-go API price on a Production key (trial keys are free but rate-limited and not for commercial use).
Capabilities
Architecture
Parameters: 111B
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Cohere docs: Cohere's Command A Model (opens in a new tab)
- Cohere release notes: Announcing Command A (opens in a new tab)
- Hugging Face: CohereLabs/c4ai-command-a-03-2025 (model card) (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Command A and Cohere
MarkTechPost
Cohere Releases Embed 5: How It Compares to Voyage 4 Large, Gemini Embedding 2, and OpenAI
Cohere has released its new embedding model family, Embed 5, designed for enterprise search, retrieval-augmented generation (RAG), and agentic retrieval applications. The model family is offered in two distinct tiers: Embed 5 Pro, which prioritizes maximum retrieval quality, and Embed 5 Fast, which is optimized for latency and cost efficiency during live query operations. A key design innovation is that both Pro and Fast tiers share a unified embedding space, allowing users to index data using one tier and query it using the other, providing significant deployment flexibility. Both tiers are capable of processing text, images, and fused text-plus-image inputs, supporting over 100 languages and handling contexts of up to 128,000 tokens. Embed 5 is available today through the Cohere API and on platforms such as Microsoft Foundry and Amazon SageMaker, with options for private VPC or on-premise serving via vLLM. The API model identifiers are embed-v5.0-pro and embed-v5.0-fast, as detailed in Cohere's model documentation. These models can output embeddings in various dimensions, including 2048 (default), 1536, 1024, 768, 512, and 256, with embeddings returned as float, int8, or binary formats. Pricing for Embed 5 Pro is set at $0.12 per 1 million text tokens, while Embed 5 Fast costs $0.08 per 1 million text tokens. Image inputs are priced at $0.40 per 1 million tokens for both tiers. The models can embed page images directly or fuse an image with its associated metadata into a single vector, a feature particularly beneficial for scanned documents, slide decks, schematics, and charts where text extraction alone may lose critical information. Cohere conducted extensive testing across 40 development datasets, evaluating various corpus and query pairings. In these tests, an index created with Embed 5 Pro and queried with Embed 5 Fast achieved a score of 98.4, normalized against a Pro-plus-Pro setup at 100. An all-Fast configuration scored 96.6. Cohere recommends a pattern of indexing with the Pro tier for optimal quality and querying with the Fast tier for speed, provided both sides utilize the same output dimension. This split is particularly advantageous for agentic workloads, where an agent might perform dozens of searches per task, making query latency a critical factor. Cohere's internal benchmarks indicate that the Fast tier processed 377.3 documents per second, compared to 159.7 documents per second for the Pro tier. On the ViDoRe V3 benchmark, Embed 5 Pro achieved an average score of 85.8, representing an 8.8-point improvement over its predecessor, Embed 4. Embed 5 Fast averaged 84.5 on the same benchmark. For comparative context, Voyage 4 Large scored 83.7, Gemini Embedding 2 scored 83.2, and OpenAI's text-embedding-3-large scored 83.0.
The Hacker News
Zyxel and Veeam Flaws Under Active Exploitation With Command and SYSTEM Access
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has added two significant vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog, indicating that both flaws are currently being actively exploited by malicious actors. The first vulnerability affects Zyxel GS1900 series network switches and is tracked as CVE-2026-7273. This flaw, which carries a CVSS score of 8.8, is a stack-based buffer overflow vulnerability that allows for arbitrary code execution. The KEV catalog entry, published on Monday, cites evidence of active exploitation, meaning that attackers have already leveraged this weakness to compromise systems. The Zyxel GS1900 series are managed switches commonly used in small to medium-sized businesses for network infrastructure management. Exploitation of this vulnerability could grant attackers significant control over the affected network devices, potentially leading to network disruption or further infiltration. In parallel, CISA also highlighted a critical vulnerability affecting Veeam Backup Enterprise Manager (VBEM), identified as CVE-2024-29849. This vulnerability, rated with a CVSS score of 9.8, is a critical authentication bypass flaw. Successful exploitation allows an unauthenticated attacker to gain administrative access to the VBEM server. This level of access is particularly concerning as VBEM is a central component for managing Veeam's backup and recovery solutions. Compromising VBEM could enable attackers to disable backups, delete backup data, or even use the compromised backup infrastructure to launch further attacks across an organization's network. The inclusion of this vulnerability in the KEV catalog underscores the immediate threat it poses to organizations relying on Veeam for data protection. CISA mandates that U.S. federal civilian executive branch agencies must apply available patches or mitigations for vulnerabilities listed in the KEV catalog by specific deadlines to protect their networks. While this mandate applies to federal agencies, the inclusion of these vulnerabilities in the KEV catalog serves as a strong warning to all organizations, regardless of sector, to prioritize patching and implementing security measures. The active exploitation of both the Zyxel and Veeam vulnerabilities suggests that attackers are actively seeking and exploiting these weaknesses to gain unauthorized access and control. Organizations using Zyxel GS1900 switches or Veeam Backup Enterprise Manager are strongly advised to review their security configurations and apply any available updates or workarounds immediately to mitigate the risk of compromise. The nature of these vulnerabilities, allowing for arbitrary code execution and administrative access, highlights the critical importance of timely patch management and robust cybersecurity practices in defending against sophisticated threats.
MarkTechPost
Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages
Cohere has released North Small Translate, an open-weight machine translation model developed by Cohere and Cohere Labs. This model is a sparse Mixture-of-Experts (MoE) architecture, featuring a total of 218 billion parameters, with 25 billion active parameters per token. North Small Translate supports translation across 50 languages, ranging from Albanian to Vietnamese. In evaluations conducted by Cohere on the WMT26 benchmark, the model achieved an average score of 83.6 across all supported languages. Cohere claims this performance surpasses that of commercial translation services like DeepL and Google Translate, as well as other open-source models such as GLM 5.2 and Mistral Large 3. The model is accessible through various deployment options: it can be used freely via Cohere's API up to specified rate limits, self-hosted for non-commercial use, or licensed for commercial applications. The development of North Small Translate represents a return to the foundational problem of machine translation, which was a key focus of the original Transformer architecture introduced by Google researchers in 2017 with the paper "Attention Is All You Need." The Transformer's initial breakthroughs were demonstrated on WMT 2014 English-to-German and English-to-French translation tasks. Cohere frames the release of this dedicated translation model as a matter of digital sovereignty, asserting that organizations unable to communicate globally may struggle to maintain their sovereignty. North Small Translate is the inaugural model within Cohere's "North" family of models and follows previous multilingual efforts such as Tiny Aya and Command A Translate. The model's real-world translation quality was shaped through collaboration with RWS, an organization whose Language Weaver scientists and language experts contributed to its development. Architecturally, North Small Translate is a decoder-only sparse MoE Transformer. It incorporates 128 experts, with 8 experts activated per token, alongside shared experts applied universally. The routing mechanism employs a sigmoid function over expert logits, normalized across the selected top-k experts. Its attention mechanism includes interleaved sliding-window layers (with a window size of 4096 and RoPE embeddings) and global layers that do not utilize positional embeddings, in a 3:1 ratio. This specific attention configuration was previously introduced in Cohere's Command A model. The model supports an input and output context window of 16,000 tokens, exclusively for text. It underwent post-training specifically to enhance translation quality. Approximately 11.5% of the model's weights are active for each token processed, and the per-token compute load corresponds to the 25 billion active parameters, although the full 218 billion parameters must be held in memory.
MarkTechPost
Cohere Releases Parse 5 (parse-v5.0): A 2.3B Vision Language Model That Turns Enterprise Documents Into Markdown
Cohere has released Parse 5 (parse-v5.0), a 2.3 billion-parameter vision language model specifically engineered for high-volume enterprise document ingestion. This model is built upon Cohere Labs’ North-Micro-Vision-Instruct architecture and features an 8,192-token context window with a footprint of approximately 4.6 gigabytes. Parse 5 is designed to process PDF, PPT, or JPEG pages, which are submitted as base64-encoded data URIs. The output is structured Markdown that includes text arranged in reading order, tables rendered in HTML format, identified lists, extracted form key-value pairs, descriptions of images, and bounding box coordinates for elements within the documents. Notably, the model does not employ a separate Optical Character Recognition (OCR) stage, integrating this functionality directly. Cohere is pricing the Parse API at $1.50 per 1,000 pages. The company positions Parse 5 on a basis of price-performance rather than solely on peak accuracy. To support this claim, Cohere reports a ParseBench score of 79.2. This score, as detailed by the company, measures three out of the five dimensions of that particular benchmark. The model is available for production deployment and is generally accessible through the Cohere Parse API, Microsoft Foundry, and AWS SageMaker, as well as via single-tenant Model Vault. There is no waitlist or research license required for access. Mid-market teams already operating a Retrieval-Augmented Generation (RAG) stack can begin using the service with metered API calls and a free trial key. Larger enterprises with specific residency or air-gap requirements can opt for direct deployment through Model Vault or private installations. While seed-stage startups can also utilize Parse 5, the economic benefits become significant for those processing upwards of approximately 100,000 pages per month. Cohere is targeting document-heavy verticals such as financial services, insurance, healthcare and life sciences, the public sector, telecommunications, energy, and manufacturing. These industries commonly deal with scanned forms and dense tables, making Parse 5 a relevant tool for their operations. Potential applications for the model include RAG ingestion, intelligent document processing, streamlining claims and invoice pipelines, enhancing contract and filing search capabilities, and providing document context to AI agents. The model's architecture, North-Micro-Vision-Instruct, is a key component enabling its vision and language processing capabilities. The 8,192-token context window allows for the processing of extensive document content in a single pass, improving efficiency for complex documents. The approximately 4.6GB model size indicates a balance between performance and deployability for enterprise environments. The integrated approach, eliminating a separate OCR step, simplifies the document processing workflow and potentially reduces latency and cost for users.
MarkTechPost
Best Open Speech Recognition (ASR) Models in 2026: WER, Languages, Latency, and License Compared
The landscape of open Automatic Speech Recognition (ASR) models has evolved beyond a single dominant player, with several models now competing on performance metrics and licensing. In March 2026, Cohere released Transcribe, a 2B parameter model under the Apache 2.0 license, which initially led the Hugging Face Open ASR Leaderboard with an average Word Error Rate (WER) of 5.42%. This was quickly followed by IBM's Granite Speech 4.1 2B, which achieved a WER of 5.33%. More recently, ARK-ASR-3B and MOSS-Transcribe-preview-2B have posted even lower WER scores, narrowing the gap at the top of the leaderboard to less than one WER point. This intense competition means that rank alone is no longer the primary factor for users selecting an ASR model. Instead, critical considerations now include the model's license, the breadth of languages it supports, its streaming capabilities for real-time processing, and the cost per audio hour. These factors are becoming increasingly important for developers and businesses integrating ASR technology into their applications. However, a closer examination of the Open ASR Leaderboard reveals nuances in how the reported average WER is calculated. The average WER cited is not a fixed quantity, and models have been scored using different test sets. For instance, Cohere's 5.42% WER is an average across eight English test sets, including TED-LIUM. In contrast, ARK-ASR-3B's 5.04% is an average across seven sets, excluding TED-LIUM. The MOSS-Transcribe-preview-2B card explicitly states that TED-LIUM is not part of its current leaderboard run. Since TED-LIUM is considered one of the easier test sets, its exclusion can artificially inflate the reported average WER. When Cohere's published per-dataset scores are recomputed using the same seven test sets that ARK reports, Cohere's average WER increases to 5.84% from the headline 5.42%. Similarly, IBM's Granite Speech 4.1 2B moves from 5.33% to 5.65% when evaluated on the same basis. This suggests that, on a like-for-like comparison, ARK's lead over other models might be larger than initially indicated by the headline leaderboard figures. The focus is shifting towards transparent and consistent evaluation methodologies to accurately compare these rapidly advancing open ASR models.