Model

Galactica 120B

OPEN

Meta · Galactica family · released Nov 15, 2022

Science LLM trained on papers/code; public demo pulled after about three days over hallucination concerns.

ReasoningCodingVisionFunction callingTool useAgentic

85.9

Elo · rank #388

Parameters

120B

Active params

120B (dense)

Context

2.048K tokens

Architecture

Dense decoder-only Transformer trained on scientific papers, code, and reference material

License

CC-BY-NC 4.0

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

ARC-ChallengeReasoning67.9%#58

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning83.8%#16

best: Phi-3-medium (14B) · 97.7%

BIG-benchReasoning48.7%#4

best: Chinchilla · 65.1%

MATH-500Math20.4%#181

best: GPT-5 · 99.4%

MMLU: AnatomyKnowledge58.5%#2

best: Palmyra Med 70B · 83.7%

MMLU: Clinical KnowledgeKnowledge59.2%#2

best: Palmyra Med 70B · 90.9%

MMLU: Medical GeneticsKnowledge68.0%#2

best: Palmyra Med 70B · 94.0%

MMLUKnowledge41.3%#256

best: OpenAI o3 · 92.9%

PubMedQAKnowledge77.6%#3

best: Palmyra Med 70B · 79.6%

TruthfulQAKnowledge26.0%#107

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

Run it locally

VRAM @ Q4

64 GB

VRAM @ FP16

240 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

llama.cpp does not support the Galactica architecture; only bitsandbytes INT8/FP16 loading is documented, and even that exceeds a single RTX 4090's VRAM.

Quantizations

bitsandbytes INT8

Galactica family

Elo progression across releases

Galactica 120BNov 202285.9

Resources

HFfacebook/galactica-120b GHpaperswithcode/galai DOCSarxiv.org/abs/2211.09085

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