command-a-plus-05-2026-bf16

by CohereLabs

218B MoE multimodal model with native citation generation for document understanding

145Kdl/month

218B total / 25B activeparams

HuggingFace Use in Pipeline

Identifiers

Model ID

CohereLabs/command-a-plus-05-2026-bf16

Feature URI

mixpeek://image_extractor@v1/cohere_command_a_plus_v1

Overview

Command A+ is Cohere's flagship open-weight multimodal model with 218B total parameters and 25B active per token. It processes scanned documents, charts, and technical manuals while generating structured output with native citations back to source material.

The model handles 128K context across 48 languages, making it particularly valuable for enterprise document understanding where provenance and citations are critical. The w4a4 quantized variant runs on a single B200 or two H100s.

Architecture

Mixture-of-Experts transformer with 218B total parameters, 25B active per token. Multimodal encoder handles interleaved text and image inputs. Native citation and grounding head produces source references alongside generated text. 128K context window.

Mixpeek SDK Integration

from mixpeek import Mixpeek
mx = Mixpeek(api_key="YOUR_KEY")

mx.ingest(
    collection_id="legal-docs",
    source="s3://contracts/",
    extractors=[
        {
            "type": "scene_caption",
            "model": "CohereLabs/command-a-plus-05-2026-bf16",
            "output_feature": "caption"
        },
        {
            "type": "text_embedding",
            "model": "BAAI/bge-m3",
            "input_field": "caption",
            "output_feature": "caption_embedding"
        }
    ]
)