Model

Tülu 3 70B

OPEN

Allen Institute for AI (Ai2) · Tülu 3 family · released Nov 21, 2024

Introduced the RLVR post-training recipe on Llama 3.1 70B; beats Llama 3.1 70B Instruct on MATH (+6.6) and BBH (+8.2)

ReasoningCodingVisionFunction callingTool useAgentic

944.0

Elo · rank #244

Parameters

70B

Active params

70B (dense)

Context

128K tokens

Architecture

Dense Transformer (Llama 3.1 70B architecture, GQA, RoPE with llama3 scaling)

License

Llama 3.1 Community License Agreement

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

AGIEvalReasoning75.0%#2

best: OLMo 3-Think 32B · 88.2%

AlpacaEvalHuman preference49.8%#20

best: Tülu 2+DPO 70B · 95.1%

BIG-Bench HardReasoning82.0%#28

best: ERNIE 4.5 300B-A47B · 94.3%

DROPReasoning74.3#39

best: Hunyuan-T1 · 93.1

GPQA DiamondReasoning48.0%#205

best: GPT-6 Astra · 96.0%

GSM8KMath93.5%#20

best: Llama 3.1 405B · 96.8%

HumanEval+Coding88.0%#7

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding92.4%#13

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning83.2%#82

best: Gemma 4 26B A4B · 98.5%

MATH-500Math63.0%#108

best: GPT-5 · 99.4%

MMLU-ProKnowledge65.8%#107

best: Claude Fable 5 · 91.5%

MMLUKnowledge83.1%#73

best: OpenAI o3 · 92.9%

TruthfulQAKnowledge67.6%#7

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

Run it locally

VRAM @ Q4

42.5 GB

VRAM @ FP16

140 GB

Fits on (Q4)

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

70B model needs ~42.5GB even at Q4_K_M — exceeds a single RTX 4090's 24GB VRAM

Quantizations

GGUF

Fine-tune it

Conditional / custom

QLoRA47.6 GB1× A100 80GB

LoRA149.1 GB1× B200 192GB

Full fine-tune1123.5 GB8× H200 141GB

QLoRA SFT on ~10k samples ≈ $49.23 (1× A100 80GB)

Plan a fine-tune →

Tülu 3 family

Elo progression across releases

Tülu 3 70BNov 2024944.0 Tülu 3 405BJan 20251160.9

Resources

HFallenai/Llama-3.1-Tulu-3-70B GHallenai/open-instruct DOCSallenai.org/olmo

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