Model
Model explorer
IDEFICS 80B
OPENHugging Face · IDEFICS family · released Aug 22, 2023
First open-access reproduction of DeepMind's Flamingo; interleaved image-text sequences via frozen CLIP-ViT-H/14 + frozen LLaMA-65B with trained cross-attention.
ReasoningCodingVisionFunction callingTool useAgentic
-264.9
Elo · unrated
Parameters
80B
Active params
80B (dense)
Context
2K tokens
Architecture
Flamingo-style decoder: frozen CLIP-ViT-H/14 vision encoder + frozen LLaMA-65B LM connected via trained Perceiver-resampler cross-attention blocks
License
Mixed: MIT (new cross-attention/connector params) + LLaMA-65B non-commercial research license (frozen backbone) + MIT (CLIP-ViT-H/14 vision encoder)
Languages
1+
API price (in/out)
No hosted API
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
best: CogVLM-17B · 148.7
best: CogVLM-17B · 94.9
best: CogVLM-17B · 64.8%
best: Molmo 2 8B · 85.7%
best: Reka Edge (2026) · 88.4%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
160 GB
Fits on (Q4)
Multi-node cluster required
Custom Flamingo-style cross-attention architecture has no mainstream llama.cpp/GGUF conversion path; 80B dense at fp16 also exceeds a single 4090 regardless.
Quantizations
—
API price weights · each benchmark row carries its own source badge (see methodology)