Model
Model explorer

OPT-66B

OPEN
Meta · OPT family · released Jun 22, 2022

Largest freely downloadable OPT checkpoint, released without gating alongside Meta's day-by-day training logbook.

ReasoningCodingVisionFunction callingTool useAgentic
-252.0
Elo · unrated
Parameters
66B
Active params
66B (dense)
Context
2.048K tokens
Architecture
Dense decoder-only Transformer (GPT-3-style architecture)
License
OPT-175B License Agreement (non-commercial research use)
Languages
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-ChallengeReasoning40.0%#149
best: Llama 3.1 405B · 96.9%
ARC-EasyReasoning67.3%#56
best: Phi-3-medium (14B) · 97.7%
LAMBADAReasoning74.9%#9
best: GPT-3 175B · 86.4%
PIQAReasoning79.8%#49
best: GPT-4o mini · 93.1%
StoryClozeReasoning77.3%#5
best: GPT-3 175B · 87.7%
Run it locally
VRAM @ Q4
35 GB
VRAM @ FP16
132 GB
Fits on (Q4)
M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
llama.cpp does not support the OPT architecture, so no RTX 4090 llama.cpp benchmark exists; the FP16 checkpoint also exceeds a single 4090's VRAM.
Quantizations
GPTQ
API price weights · each benchmark row carries its own source badge (see methodology)