Model

OLMo 3-Think 32B

OPEN

Allen Institute for AI (Ai2) · OLMo 3 family · released Nov 20, 2025

Latest flagship and strongest fully open reasoning model; narrows the gap to Qwen 3-32B-Thinking on reasoning benchmarks while trained on ~6x fewer tokens.

ReasoningCodingVisionFunction callingTool useAgentic

1457.1

Elo · rank #130

Parameters

32.23B

Active params

32.23B (dense)

Context

64K tokens

Architecture

Dense transformer (decoder-only, sliding/full attention mix, exposed CoT)

License

Apache 2.0

Languages

API price (in/out)

$0.15 / $0.5

Modalities

text

Benchmark results

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

AGIEvalReasoning88.2%#1

best: this model · 88.2%

AIMEMath76.8%#88

best: GPT-5.2 · 100.0%

AlpacaEvalHuman preference74.2%#7

best: Tülu 2+DPO 70B · 95.1%

BIG-Bench HardReasoning89.8%#6

best: ERNIE 4.5 300B-A47B · 94.3%

GPQA DiamondReasoning61.0%#162

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning5.9%#112

best: Claude Sonnet 5 · 57.4%

HumanEval+Coding91.4%#6

best: Mistral Small 3.2 (24B) · 92.9%

IFBenchReasoning47.6%#38

best: MiniMax M3 · 83.0%

IFEvalReasoning89.0%#29

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding83.5%#33

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math96.1%#28

best: GPT-5 · 99.4%

MBPP+Coding68.0%#19

best: Llama 3.1 405B · 88.6%

MMLU-ProKnowledge75.9%#72

best: Claude Fable 5 · 91.5%

MMLUKnowledge85.4%#56

best: OpenAI o3 · 92.9%

Run it locally

VRAM @ Q4

20 GB

VRAM @ FP16

65 GB

Fits on (Q4)

RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF Q4

OLMo 3 family

Elo progression across releases

OLMo 3 7BNov 2025892.8 OLMo 3-Think 32BNov 20251457.1

Resources

HFallenai/Olmo-3-32B-Think GHallenai/OLMo-core DOCSallenai.org/blog/olmo3 APIAvailable via Allen Institute for AI (Ai2) API

API price $0.15/$0.5 · each benchmark row carries its own source badge (see methodology)