Model
Model explorer

Idefics2 8B

OPEN
Hugging Face · IDEFICS family · released Apr 15, 2024

Compact Apache-2.0 VLM roughly 10x smaller than IDEFICS-80B, with strong OCR/document QA via SigLIP + Mistral-7B.

ReasoningCodingVisionFunction callingTool useAgentic
564.3
Elo · unrated
Parameters
8B
Active params
8B (dense)
Context
32K tokens
Architecture
SigLIP-SO400M vision encoder + Mistral-7B-v0.1 decoder, connected via learned Perceiver resampler + MLP modality projection
License
Apache-2.0
Languages
—
API price (in/out)
No hosted API
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
DocVQAVision74.0%#43
best: Qwen2-VL-72B · 96.5%
MathVistaVision51.4%#52
best: Seed 2.1 Pro · 90.7%
MMBenchVision76.7%#4
best: Phi-3.5-vision (4.2B) · 81.9%
MMMUVision43.0%#102
best: Claude Fable 5 · 89.3%
TextVQAVision73.0%#20
best: Molmo 2 8B · 85.7%
VQAv2Vision81.2%#5
best: Reka Edge (2026) · 88.4%
Run it locally
VRAM @ Q4
6 GB
VRAM @ FP16
17 GB
Fits on (Q4)
RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
llama.cpp only converts the Mistral-7B LLM portion (mmproj vision tower unsupported), so no verified end-to-end Q4 4090 throughput exists.
Quantizations
AWQ
IDEFICS family
Elo progression across releases
IDEFICS 80BAug 2023-264.9Idefics2 8BApr 2024564.3
Resources
HFHuggingFaceM4/idefics2-8bDOCShuggingface.co/blog/idefics2
API price weights · each benchmark row carries its own source badge (see methodology)