Cohere: Command R+ (08-2024)

Processes multi-turn conversations with built-in support for retrieval-augmented generation (RAG) through Cohere's native document grounding API. The model maintains conversation context across turns while integrating external document retrieval, enabling it to cite sources and ground responses in provided documents without requiring manual prompt engineering for RAG patterns.

Implements function calling through JSON schema-based tool definitions, allowing the model to decide when and how to invoke external APIs or functions. The model generates structured tool calls with parameters that conform to provided schemas, enabling agentic workflows where the model orchestrates multiple tools across reasoning steps without explicit prompt templates.

Processes documents and conversations up to 128K tokens using optimized attention mechanisms (likely sliding window or sparse attention patterns) that reduce computational complexity from O(n²) to near-linear scaling. This enables processing of entire books, codebases, or conversation histories without truncation while maintaining sub-second latency through the 08-2024 performance optimization (25% lower latency vs previous version).

Extracts structured information from unstructured text by constraining generation to conform to provided JSON schemas, ensuring output always matches expected data structures. The model generates valid JSON that adheres to field types, required properties, and nested object structures without post-processing or validation failures, enabling reliable ETL pipelines and data enrichment workflows.

Ranks and retrieves relevant documents from collections based on semantic similarity to queries, using dense vector embeddings computed by the model's encoder. The ranking mechanism considers both semantic relevance and document metadata, enabling hybrid search that combines keyword and semantic signals without requiring separate embedding models or vector databases.

Generates and understands text across 100+ languages with shared embedding space enabling cross-lingual transfer — a query in English can retrieve documents in Spanish, and responses can be generated in the user's language without language-specific fine-tuning. The model uses a unified tokenizer and embedding space trained on multilingual corpora, enabling zero-shot language switching within conversations.

Follows detailed, multi-step instructions with high fidelity by decomposing complex tasks into intermediate reasoning steps and validating outputs against instruction constraints. The model maintains instruction context across long sequences and handles edge cases specified in instructions without requiring explicit prompt engineering for each variation, using chain-of-thought-like reasoning patterns internally.

Manages multi-turn conversations with automatic context optimization that selectively retains relevant information across turns while pruning redundant or outdated context. The model tracks conversation state implicitly and can reference earlier turns without explicit context passing, using attention mechanisms to weight recent and relevant turns more heavily than distant turns.

Generates responses that avoid harmful, toxic, or unsafe content through alignment training and built-in safety classifiers that detect and mitigate unsafe outputs. The model refuses requests for illegal activities, violence, or hate speech, and provides explanations for refusals rather than silent filtering, enabling transparent safety boundaries in production systems.

Processes large batches of requests (100s-1000s) with optimized throughput that leverages the 08-2024 performance improvements (50% higher throughput). Batching is handled transparently by the API, allowing callers to submit multiple independent requests that are processed in parallel on shared GPU resources, reducing per-request latency through amortized overhead.