Model

OPT-175B

OPEN

Meta · OPT family · released May 3, 2022

First GPT-3-scale (175B) open-weight model, released with training code and a full day-by-day training logbook.

ReasoningCodingVisionFunction callingTool useAgentic

-138.8

Elo · unrated

Parameters

175B

Active params

175B (dense)

Context

2.048K tokens

Architecture

Dense decoder-only Transformer (GPT-3-style architecture)

License

OPT-175B License Agreement (non-commercial research use, gated access)

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning43.9%#141

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning71.0%#47

best: Phi-3-medium (14B) · 97.7%

HellaSwagReasoning59.3%#141

best: Claude 3 Opus · 95.4%

LAMBADAReasoning75.6%#8

best: GPT-3 175B · 86.4%

OpenBookQAReasoning34.0%#48

best: Claude 1 · 90.8%

PIQAReasoning81.1%#35

best: GPT-4o mini · 93.1%

StoryClozeReasoning79.8%#4

best: GPT-3 175B · 87.7%

WinoGrandeReasoning72.6%#85

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

90 GB

VRAM @ FP16

350 GB

Fits on (Q4)

M3 Ultra 512GBH200 141GBB200 192GB

llama.cpp does not support the OPT architecture and the model far exceeds a single RTX 4090's 24GB VRAM even at 4-bit; no such benchmark exists.

Quantizations

GPTQ

OPT family

Elo progression across releases

OPT-175BMay 2022-138.8 OPT-66BJun 2022-252.0 OPT-IML 30BDec 2022-189.2

Resources

HFfacebook/opt-175b GHfacebookresearch/metaseq DOCSarxiv.org/abs/2205.01068

API price weights · each benchmark row carries its own source badge (see methodology)