Atlas

Models

← All models

Command R7B (Dec 2024)

Cohere · 2024-12-13 · 8B parameters

Cohere Labs Command R7B is a December 2024 open-weight multilingual model from Cohere and Cohere Labs with exactly 8,028,033,024 checkpoint parameters, marketed as 7B, and aligned through supervised fine-tuning and preference training for retrieval-augmented generation, multi-step tool use, reasoning, summarization, question answering, and code across 23 languages. Its 32-layer decoder-only Transformer interleaves three 4,096-token sliding-window grouped-query-attention layers using RoPE for each global-attention layer without positional embeddings. The current checkpoint configuration declares 132,096 maximum positions, while Cohere’s model card and hosted documentation document a 128,000-token context window.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)55.0
BFCL v4 Overall Accuracy32.1
BuseyBench SVG1.2
Capability112.9
Global PIQA - Non-Parallel (Strict Exact Match)59.0
Global PIQA - Parallel (Strict Exact Match)33.0
GPQA (Open LLM Leaderboard v2)30.9
IFEval (Open LLM Leaderboard v2)77.1
MATH Level 5 (Open LLM Leaderboard v2)29.9
MMLU-Pro (Open LLM Leaderboard v2)35.7
MuSR (Open LLM Leaderboard v2)41.3
SnakeBench Average Apples0.3
SnakeBench Best Apples1.0
SnakeBench Rating13.5
SnakeBench Total Apples9.0
SnakeBench Win Rate34.5
Loading Atlas data…