Model

UL2 20B

OPEN

Google · UL2 family · released Oct 14, 2022

20B open encoder-decoder trained with Mixture-of-Denoisers; strong chain-of-thought despite small size.

ReasoningCodingVisionFunction callingTool useAgentic

376.5

Elo · rank #338

Parameters

20B

Active params

20B (dense)

Context

0.5K tokens

Architecture

Dense encoder-decoder Transformer (20B, Mixture-of-Denoisers)

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

GSM8KMath6.9%#182

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning94.1%#3

best: Claude 3 Opus · 95.4%

MMLUKnowledge39.2%#258

best: OpenAI o3 · 92.9%

OpenBookQAReasoning87.2%#5

best: Claude 1 · 90.8%

PIQAReasoning90.7%#2

best: GPT-4o mini · 93.1%

SuperGLUEReasoning90.7%#1

best: this model · 90.7%

WinoGrandeReasoning90.1%#2

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

40 GB

Fits on (Q4)

Multi-node cluster required

No confirmed GGUF/quantized builds found; encoder-decoder architecture is not a standard llama.cpp target, so no measured RTX 4090 figures exist.

Quantizations

—

UL2 family

Elo progression across releases

UL2 20BOct 2022376.5

Resources

HFgoogle/ul2 GHgoogle-research/google-research DOCSarxiv.org/abs/2205.05131

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