Cohere Command R (08-2024)

Despre Cohere

Fondat 2019 · Toronto, Canada

Cohere is an enterprise-focused LLM lab founded in 2019 by Aidan Gomez (co-author of 'Attention Is All You Need'), Ivan Zhang and Nick Frosst. Headquartered in Toronto with offices in San Francisco and London, Cohere targets enterprise customers — banks, telcos, government — with deployment options spanning Cohere's hosted API, AWS Bedrock, Azure, Oracle Cloud and on-prem (Cohere North). The Command R family launched March 2024 as Cohere's RAG-and-tool-use flagship line, refreshed as command-r-08-2024 in August 2024 with substantial gains on math, code, reasoning and reduced spurious refusals. Investors include Nvidia, Cisco, Salesforce and Oracle, with total funding above $1B and a 2024 valuation around $5.5B.

Arhitectură

Decoder-only Transformer optimised for RAG and tool use

Command R 08-2024 is a 35B dense decoder-only transformer designed specifically for retrieval-augmented generation and multi-step tool use. The architecture uses grouped-query attention (8 KV heads), RoPE positional embeddings (extended for 128K context), SwiGLU activations, RMSNorm and no biases, with a 256,000-token multilingual BPE tokeniser shared with the Aya line. The 08-2024 refresh kept the architecture and tokeniser identical to the March 2024 release but substantially improved math, code, reasoning, structured output and reduced 'I cannot answer' refusals. Post-training emphasises grounded citation generation (Cohere's structured `<co: 0>` citation tokens) and parallel multi-step tool calling. Open weights are released under CC-BY-NC 4.0 for research; commercial use goes through the hosted Cohere API, AWS Bedrock, Azure, Oracle marketplaces or licensed on-prem deployment.

Parametri

35B (dense)

Context

128.000 tokeni

Capabilități

• Excellent retrieval-augmented generation with structured inline citations
• Native multi-step tool/function calling with parallel-tool support
• 128K context window for long-document analysis
• Strong multilingual quality across 10+ business languages
• Improved math, code and reasoning in the 08-2024 refresh
• Open weights (CC-BY-NC) for research evaluation
• Hosted on Cohere API, AWS Bedrock, Azure, Oracle Cloud
• Best for: enterprise RAG, multilingual support agents, mid-tier tool-using assistants.

Antrenament & licență

Trained on trillions of tokens of multilingual web data, code, math and Cohere-curated synthetic RAG and tool-use traces. Knowledge cutoff approximately early 2024. Post-training is supervised fine-tuning followed by preference optimisation, with specific data and methods proprietary to Cohere.

Licență: Cohere Terms of Service for the hosted API. Open weights on Hugging Face under CC-BY-NC 4.0 (research-only). Commercial production access via Cohere API, AWS Bedrock, Azure, Oracle Cloud or licensed on-prem deployments.

Teste de siguranță: Internal Cohere safety evaluation aligned with the Cohere Acceptable Use Policy. Cohere publishes model cards and an enterprise-focused safety profile, including reduced over-refusal in the 08-2024 refresh.