Atlas

Models

← All models

C4AI Command R+

Cohere · 2024-04-04 · 103.8B parameters

C4AI Command R+ is Cohere and Cohere For AI's original 104B open-weight instruction-tuned checkpoint, released on April 4, 2024 and post-trained with supervised fine-tuning and preference training for retrieval-augmented generation, grounded citations, multilingual generation, and multi-step tool use. Its model card advertises a 128K context, while the released configuration sets model_max_length to 131,072 alongside max_position_embeddings of 8,192; it is distinct from the August 2024 refresh.

Benchmark scores

BenchmarkScore
Agentic Tool Use (Chat)20.2
Agentic Tool Use (Enterprise)30.2
Artificial Analysis Intelligence Index3.0
BBH (Open LLM Leaderboard v2)58.2
BigCodeBench Complete41.9
BigCodeBench Instruct33.8
BigCodeBench-Hard Complete14.2
BigCodeBench-Hard Instruct13.5
Capability111.0
Creative Writing v264.8
EQ-Bench v276.1
EQ-Bench v2 + MAGI-Hard Combined62.9
GPQA (Open LLM Leaderboard v2)30.5
GPQA Diamond28.3
GPQA Diamond (Vals protocol average)31.1
Humanity's Last Exam (Text-Only)4.5
IFEval (Open LLM Leaderboard v2)76.6
LegalBench68.3
LegalBench - Conclusion Tasks48.3
LegalBench - Interpretation Tasks71.6
LegalBench - Issue Tasks76.3
LegalBench - Rhetoric Tasks75.6
LegalBench - Rule Tasks69.6
LiveCodeBench (Artificial Analysis)12.2
LiveCodeBench (Vals AI public dataset)18.2
LiveCodeBench (Vals Public) - Easy48.4
LiveCodeBench (Vals Public) - Hard0.0
LiveCodeBench (Vals Public) - Medium6.3
MAGI-Hard49.7
MATH Level 5 (Open LLM Leaderboard v2)8.0
Loading Atlas data…