Cohere targets global enterprises with new highly multilingual Command A model requiring only 2 GPUs

Canadian AI Start Coune – withdrew by one of the authors of the original transformer paper This is a large language model (LLM) revolution in 2017 – started today an open command aThe latest generative AI model designed for enterprise applications.

Like the heir Team-r who debut in March 2024and After R + command The command commander is aimed at the efficiency of the use of the search widened generation (dwarf), external instrument use and enterprise AI – especially for computing and responding speed.

The bank intends to make an attractive choice for businesses who want to get an advantage of the AI ​​advantage for applications without breaking and emergency answers – finance, health, medicine, science and law.

The faster speed, low hardware requirements and multilingual opportunities, for example, to order a position command to models as a powerful alternative GPT-4O and Deepseek-v3 – Recently, new justification models that take the AI ​​industry by the storm, classic LLMS.

A context length of 128,000 token, which supports a context length (can be connected to an entry / exit exchange of LLC, a 300-page romaniary, 256,000 Token (600 pages equivalent).

The AUs also come to the heights coager – a non-profit subsidiary company – a multilingual vision model called open source (for research only) This month before the moon Vision.

One step from the team-r

When the team-r begins in early 2024, he presented key innovations such as optimized dwarf performance, better knowledge search and low-precious AI placements.

Oracle, concept, scale, AI, AI, AI, integrated into business solutions such as McKinsey A November 2024 Report from Menlo Enterprise Research Enterprise Enterprise Place the market share of the area between 3%, open (34%), anthropic (24%), an anthropic (24%), even a smaller beginning.

Now, in the offer to have a bigger enterprise, the order puts forward these opportunities. According to the colare, this is:

• Matches or Outperforms Openai’s GPT-4O and DeepSeek-V3, work, root and coding positions
• Only two GPUs (A100 or H100) work with improving great efficiency than models required to 32 GPUs
• GPT-4O and 2.4x faster than 156 verses – faster than 1.4x and 2.4x faster than 1.4x faster
• The delay with Token, which is 6.500mms time compared to 7,460 s for GPT-4O, is reduced, DeepSeek-v3
• Enhances multilingual AI capabilities, with an Arabic dialect and expanded support adapted for 43 global languages.

Not in it Developer documents online “There is no community of command.

Built for enterprise

COUNE continued its first strategy for the first strategy, and it has provided a problem with a problem with the work environment. The main features include:

• Advanced Search Extended Generation (Dwarf): Provides checkable, high-precision answers for enterprise applications
• Use Agent Vehicle: Integrating with enterprise tools, supports complex workflows
• Integration of the North AI platform: Works with Coone’s North AI platform, allowing institutions to automate tasks using safe, enterprise-grade AI agents
• Death ability and cost efficiency: Personal Placement is 50% cheaper than API-based entrance.

Multilingual and high performers in Arabic

The commanding feature of the command is the ability to create accurate answers between 23 of the world’s most common language, including improved operating processing. Supported languages ​​(according to) Development documents on the CoE Website) These are:

• English
• French
• Spanish
• Italian
• German
• Portuguese
• Japanese
• Korean
• Chinese
• Arab
• Russian
• Polish
• Turkish
• Vietnamese
• Dutch
• Czech
• Indonesia
• Ukrainian
• Romanian
• Greek
• Hindi
• Hebrew
• Persian

In evaluation assessments:

• The order consisting of 98.2% in response to Arabic, is higher in English – DeepSeek-V3 (94.9%) and GPT-4O (92.2%).
• Compared to 15.9 (GPT-4O) and 15.7 (GPT-4O) and 15.7 (GPT-4O) and 15.7 (DPSeek-V3), the name of the name 2 in the account of 24.7 accounts exceeds the rivals.

Established for speed and efficiency

Speed ​​is a critical factor for the placement of AI and engineered to give faster results than many opponents.

• Token stream speed for 100k context: 73 Tokens / Sec (38 / second compared to GPT-4O and DeepSek-v3 in 32// seconds)
• Faster first trait generation: The response period significantly reduces other large-scale models compared to other large-scale models

Price and availability

Command aer already available Cohere platform and Open Weights for Research Use Only Hugging face a Creative Commons Attribution Non-Commercial 4.0 International (CC-by-NC 4.0) LicenseSupport for a wider cloud provider will come soon.

• Input Tokens: $ 2.50 per million
• Output Tokens: $ 10.00 for Million

Available with personal and on-Prem deployment request.

Industrial reactions

Several AI researchers and coone team members shared their enthusiasm for the team.

Consaknath Ganesan, on the contrary, said: “The command is amazing to be placed in the last few months. The command can be placed in 2 H100 GPU! 256K Contextual support, Extended multilingual support, agent tool … I am very proud.”

Pierre Richemond, EU researcher, “Command A Command A new GPT-4O / DeepSEEK V3 is a 256k context length optimized for the effectiveness of the enterprise, open weights, enterprises.”

The creation of the team-R Foundation, the COONE team represents the next step in the expanding, efficient enterprise AI.

The faster speed, multilingual handling and low placement costs with a larger context window have been improved, providing powerful alternative businesses to existing AI models.