Model
Model explorer

Tülu 2+DPO 70B

OPEN
Allen Institute for AI (Ai2) · Tülu 2 family · released Nov 17, 2023

Zephyr-style DPO recipe applied to Llama-2-70B SFT; largest publicly released DPO-trained model at the time

ReasoningCodingVisionFunction callingTool useAgentic
523.6
Elo · rank #297
Parameters
69B
Active params
69B (dense)
Context
8K tokens
Architecture
Dense Transformer (Llama 2 70B architecture, GQA, RoPE, RMSNorm)
License
AI2 ImpACT Low-risk License (built on Llama 2 Community License)
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AlpacaEvalHuman preference95.1%#1
best: this model · 95.1%
BIG-Bench HardReasoning66.0%#70
best: ERNIE 4.5 300B-A47B · 94.3%
GSM8KMath71.5%#99
best: Llama 3.1 405B · 96.8%
MMLUKnowledge67.8%#162
best: OpenAI o3 · 92.9%
MT-BenchHuman preference7.89#49
best: Hunyuan-Large (A52B) · 9.4
Run it locally
VRAM @ Q4
41.4 GB
VRAM @ FP16
138 GB
Fits on (Q4)
M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
70B model needs ~41GB even at Q4_K_M — exceeds a single RTX 4090's 24GB VRAM
Quantizations
GGUF · AWQ · GPTQ
Tülu 2 family
Elo progression across releases
Tülu 2+DPO 70BNov 2023523.6
Resources
HFallenai/tulu-2-dpo-70bGHallenai/open-instructDOCSallenai.org/impact-license
API price weights · each benchmark row carries its own source badge (see methodology)