Cohere's smallest vision-language model to date: 2.4B parameters, open weights under Apache 2.0, with native-resolution image input that preserves aspect ratio instead of squashing every input into a small square. It pairs a custom-trained 400M-parameter native-resolution vision encoder with an in-house 2B language model, North Micro LLM, which follows the Command A+ architecture — three sliding-window attention layers using rotary embeddings interleaved with one global attention layer carrying no positional embeddings (the released config.json declares model type cohere_compass, 28 layers, hidden size 2048, 16 heads with 8 KV heads and a 4,096-token sliding window). The encoder was continued pre-trained from Google's SigLIP 2 SO400M checkpoint and given C-RoPE — 2D RoPE plus bilinearly interpolated learned 1D position embeddings — to support native-resolution inputs; the projector follows DeepStack, injecting patch embeddings from several encoder layers into the corresponding early LLM layers. Tokenizer vocabulary is 262,144 and the checkpoint is bfloat16. The language backbone takes 128K tokens, but multimodal training used an 8K-token context and Cohere does not claim validation beyond it.

Four training stages are documented in the technical blog. Stage 1 (10M examples) trained only the encoder and projector at a fixed 384×384 with the LM frozen, on 60% dense captions and 40% OCR. Stage 2.1 (13M) raised resolution to 1024×1024 and trained encoder, projector and LM jointly on a 50/50 mix. Stage 2.2 (10M) raised the native-resolution cap to 1654×2339 pixels — an A4 page at 200 dpi. Stage 3 (50M) instruction-tuned the full model on a mixture led by native OCR and charts/tables (17.8% each), grounding/counting and OCR QA (13.3% each) and general VQA (11.2%), keeping 8.9% text-only data to preserve language ability. Stage 4 (500k) applied a simplified Mixed Preference Optimization: Cohere dropped MPO's BCO quality term and combined DPO with an auxiliary next-token loss weighted at 15%, training only the LM. Next-token prediction was the primary objective throughout, with a patch-token alignment auxiliary loss at 0.05 weight. Training data drew on public datasets plus an in-house multilingual document corpus used to synthesize OCR, document, chart, captioning, HTML-table and grounding examples; bounding boxes are normalized to [x1,y1,x2,y2] on a 0–1000 scale. Cohere's own data ablations report that once the lowest-scoring examples were filtered out, more aggressive thresholds and small mixture changes had only marginal effect — coverage and balance mattered more than tuning proportions.

Evaluated with VLMEvalKit against seven compact open-weight models with generation capped at 1,024 tokens, the profile is document- and grounding-heavy: DocVQA (VAL) 0.921, ChartQA (Test) 0.808, OCRBench 0.792, AI2D (TEST) 0.775 and RefCOCO (avg) 0.732 — the grounding figure against 0.317 for Ministral-3-3B-Instruct and 0.304 for Qwen3-VL-2B-Instruct, though Qwen3.5-2B-Instruct leads at 0.785. It is distinctly weaker on STEM and text-only work: MMMU (DEV_VAL) 0.329 against 0.508 for Ministral-3-3B and 0.477 for Gemma-4-E2B-it, MMLU 0.504 and MMLU-Pro 0.307. Cohere states the model is not a reasoning model, has limited math and code generation, does not support tool calling or agentic workflows, and was not trained with system prompts. Public vLLM support was listed as coming soon at release; NVIDIA ships an AutoModel fine-tuning recipe, and Axolotl and MLX-VLM support are community-contributed. The date marks the HuggingFace repository's creation; the technical blog, by David Rau and Florian Schneider of Cohere Labs, was published 2026-08-12. The weights sit in the CohereLabs HuggingFace org, and the vision encoder's SigLIP 2 origin is the only external checkpoint in the stack — the language backbone is Cohere's own.

Model Details

Architecture DENSE
Parameters 2.4B
Context window 128,000
License Apache 2.0

Benchmark Scores

Benchmark Score Mode
DocVQA (VAL) 92.1% —
ChartQA (Test) 80.8% —
OCRBench 79.2% —
AI2D (TEST) 77.5% —
IFEval 74.9% —
RefCOCO (avg) 73.2% —
InfoVQA (VAL) 65.2% —
MMBench (DEV_EN_V11) 68.7% —
MMStar 51.8% —
MMMU (DEV_VAL) 32.9% —
MMLU 50.4% —
visionvision-languagemultimodalopen-weightsmall-modelmultilingualdocument-aiocr

Related