Vision: SigLIP ViT-SO400M (27 layers, 1152d, 384px, multi-layer features [-24,-20,-12,-1])
Projector: 2-layer MLP (4×1152 → 2048)
LLM: Granite-3.1-2B (40 layers, 2048d, GQA 32/8)
- embedding_multiplier=12.0, residual_multiplier=0.22, logits_scaling=8.0

Models

File	Quant	Size	Notes
`granite-vision-3.3-2b-f16.gguf`	F16	5.6 GB	Full precision
`granite-vision-3.3-2b-q8_0.gguf`	Q8_0	3.2 GB	Best quality/size
`granite-vision-3.3-2b-q4_k.gguf`	Q4_K	1.9 GB	Smallest (vision stays F16)
`granite-vision-ref.gguf`	F32	36 MB	Reference activations for parity

Parity

Vision encoder + projector verified via crispembed-diff:

crispembed --ocr granite-vision-3.3-2b-q8_0.gguf document.png

Upstream model: ibm-granite/granite-vision-3.3-2b — published by ibm-granite.
Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.