Model

Chinchilla

CLOSED

Google · Chinchilla family · released Mar 29, 2022

70B model trained compute-optimally on 1.4T tokens; outperformed 280B Gopher, reshaping LLM scaling laws.

ReasoningCodingVisionFunction callingTool useAgentic

310.8

Elo · rank #345

Parameters

70B

Active params

70B (dense)

Context

2K tokens

Architecture

Dense decoder-only Transformer (70B, compute-optimal)

License

Proprietary (Google DeepMind; research model, weights never released)

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

BIG-benchReasoning65.1%#1

best: this model · 65.1%

HellaSwagReasoning80.8%#74

best: Claude 3 Opus · 95.4%

LAMBADAReasoning77.4%#6

best: GPT-3 175B · 86.4%

MMLUKnowledge67.6%#163

best: OpenAI o3 · 92.9%

PIQAReasoning81.8%#29

best: GPT-4o mini · 93.1%

RACE-HReasoning82.3%#7

best: Claude 3 Opus · 92.9%

Social IQaReasoning51.3%#22

best: Apple DCLM-Baseline 7B · 82.9%

TriviaQAKnowledge73.2%#19

best: Sarvam-1 (2B) · 90.6%

TruthfulQAKnowledge43.6%#70

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

WinoGrandeReasoning74.9%#67

best: PaLM 2 · 90.9%

Run it locally

Closed weights — available via API only. No local deployment.

Input / M tok

—

Output / M tok

—

Chinchilla family

Elo progression across releases

ChinchillaMar 2022310.8

Resources

DOCSarxiv.org/abs/2203.15556

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